genetic_POPULATIONS.doc

834
J Glob Optim (2007) 37:405–436

Transcript of genetic_POPULATIONS.doc

Page 1: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

Page 2: genetic_POPULATIONS.doc

DOI 10.1007/s10898-006-9056-6

Page 3: genetic_POPULATIONS.doc

O R I G I N A L P A P E R

Page 4: genetic_POPULATIONS.doc

On initial populations of a genetic algorithm

Page 5: genetic_POPULATIONS.doc

for continuous optimization problems

Page 6: genetic_POPULATIONS.doc

HeikkiMaaranen · KaisaMiettinen · Antti Penttinen

Page 7: genetic_POPULATIONS.doc

Received: 7 June 2006 / Accepted: 10 June 2006 /

Page 8: genetic_POPULATIONS.doc

Published online: 20 July 2006

Page 9: genetic_POPULATIONS.doc

© Springer Science+Business Media B.V. 2006

Page 10: genetic_POPULATIONS.doc

Abstract Genetic algorithms are commonly used metaheuristics for global optimi- zation, but there has been very little research done on the generation of their initial population. In this paper, we look for an answer to the question whether the initial population plays a role in the performance of genetic algorithms and if so, how it

Page 11: genetic_POPULATIONS.doc

should be generated. We show with a simple example that initial populations may

Page 12: genetic_POPULATIONS.doc

have an effect on the best objective function value found for several generations. Traditionally, initial populations are generated using pseudo random numbers, but

Page 13: genetic_POPULATIONS.doc

there are many alternative ways. We study the properties of different point genera- tors using four main criteria: the uniform coverage and the genetic diversity of the

Page 14: genetic_POPULATIONS.doc

points as well as the speed and the usability of the generator. We use the point gen-

Page 15: genetic_POPULATIONS.doc

erators to generate initial populations for a genetic algorithm and study what effects

Page 16: genetic_POPULATIONS.doc

the uniform coverage and the genetic diversity have on the convergence and on the final objective function values. For our tests, we have selected one pseudo and one quasi random sequence generator and two spatial point processes: simple sequential

Page 17: genetic_POPULATIONS.doc

inhibition process and nonaligned systematic sampling. In numerical experiments, we solve a set of 52 continuous test functions from 16 different function families, and

Page 18: genetic_POPULATIONS.doc

analyze and discuss the results.

Page 19: genetic_POPULATIONS.doc

Keywords Global optimization · Continuous variables · Evolutionary algorithms ·

Page 20: genetic_POPULATIONS.doc

Initial population · Random number generation

Page 21: genetic_POPULATIONS.doc

H. Maaranen

Page 22: genetic_POPULATIONS.doc

Patria Aviation Oy, Lentokonetehtaantie 3, FI-35600 Halli, Finland

Page 23: genetic_POPULATIONS.doc

K. Miettinen (B)

Page 24: genetic_POPULATIONS.doc

Helsinki School of Economics, P.O. Box 1210, FI-00101 Helsinki, Finland

Page 26: genetic_POPULATIONS.doc

A. Penttinen

Page 27: genetic_POPULATIONS.doc

Department of Mathematics and Statistics,

Page 28: genetic_POPULATIONS.doc

P.O. Box 35 (MaD), FI-40014 University of Jyväskylä, Finland

Page 29: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 406

Page 30: genetic_POPULATIONS.doc

1 Introduction

Page 31: genetic_POPULATIONS.doc

When solving real optimization problems numerically, the solution process typically

Page 32: genetic_POPULATIONS.doc

involves the phases of modeling, simulation and optimization. A simulated model of a real life problem is often complex, and the objective function to be minimized may

Page 33: genetic_POPULATIONS.doc

be nonconvex and have several local minima. Then, global optimization methods are needed to prevent the stagnation to a local minimum. Therefore, in the recent years, there has been a great deal of interest in developing methods for solving global opti- mization problems (see, e.g., [12, 18, 32] and references therein). Here, we consider

Page 34: genetic_POPULATIONS.doc

global continuous optimization problems.

Page 35: genetic_POPULATIONS.doc

Genetic algorithms [15, 16, 25] are metaheuristics used for solving problems with

Page 36: genetic_POPULATIONS.doc

both discrete and continuous variables. The population is the main element of genetic

Page 37: genetic_POPULATIONS.doc

algorithms, and the genetic operations like crossover and mutation are just instru- ments for manipulating the population so that it evolves towards the final population

Page 38: genetic_POPULATIONS.doc

including a “close to optimal” solution. The requirements set on the population also

Page 39: genetic_POPULATIONS.doc

change during the execution of the algorithm.

Page 40: genetic_POPULATIONS.doc

In the recent years, genetic operators have been developed intensively (see, for

Page 41: genetic_POPULATIONS.doc

example, [25, 26] and references therein). In addition, most of the theoretical studies involve the tuning or controlling of the parameters of genetic operators and their role

Page 42: genetic_POPULATIONS.doc

is often considered significantperformance-wise (see, forexample, [11]).However, the role of the initial population, which is the topic of this paper, is widely ignored. Often, the whole area of research is set aside by a statement “generate an initial population,” without implying how it should be done. We show with a simple example that initial

Page 43: genetic_POPULATIONS.doc

populations may have effects, on the best objective function value found, and these effects may last for several generations. Then, we continue to study whether the tra- ditional way of generating initial populations is recommendable or whether there are other point sets that give faster convergence. Our motivation is to encourage discus-

Page 44: genetic_POPULATIONS.doc

sion onwhether one should paymore attention to the generation of initial populations.

Page 45: genetic_POPULATIONS.doc

We concentrate on the casewhere there is no a priori information about the location

Page 46: genetic_POPULATIONS.doc

of the global minima. Then, initial populations of genetic algorithms are traditionally

Page 47: genetic_POPULATIONS.doc

generated“randomly.” In practice, genuinerandom (truly independent) points cannot be generated numerically, and instead, pseudo random points (see, for example, [14]) are used, which imitate genuine random points. However, beside [23, 24], there is, up to our knowledge, practically no research done on whether the initial population

Page 48: genetic_POPULATIONS.doc

should be random.

Page 49: genetic_POPULATIONS.doc

In [23, 24], quasi random sequences are used to generate initial populations for ge-

Page 50: genetic_POPULATIONS.doc

netic algorithms. Quasi random points do not imitate random points but are designed

Page 51: genetic_POPULATIONS.doc

to maximally avoid each other [33]. Quasi random sequences are used in numerical

Page 52: genetic_POPULATIONS.doc

integration [5, 35, 38], computer simulations [1, 22] and quasi random searches [21, 29, 37] with good success. For example, in [2, 3], modified controlled random searches

Page 53: genetic_POPULATIONS.doc

and topographical multilevel single linkage are proposed, respectively, which use

Page 54: genetic_POPULATIONS.doc

quasi random sequences when generating an initial set of solutions. The comparison shows that the proposed methods perform significantly better than the other algo- rithms in the comparison [2, 3]. However, in [2, 3], the influence of the use of quasi random sequences alone cannot be estimated since also other modifications are made

Page 55: genetic_POPULATIONS.doc

simultaneously to the algorithms.

Page 56: genetic_POPULATIONS.doc

In this paper, we single out the effect of the different initial population for genetic

Page 57: genetic_POPULATIONS.doc

algorithms by keeping the rest ofalgorithm identical.We study the influence of the ini- tial population more generally than in [23, 24] and discuss the properties of different

Page 58: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

l u ∈Rn

407

i i i ,

i

Page 59: genetic_POPULATIONS.doc

types of point generators and the effects that different initial populations have on the convergence and the final objective function values. We also collect information and references about different ways of generating initial populations for those who are

Page 60: genetic_POPULATIONS.doc

interested in alternativeways. For the convenience of the reader,we briefly summarize ways of generating points not widely used in the field ofoptimization. In thenumerical tests, we use a well-established pseudo random number generator, a so-called Nie-

Page 61: genetic_POPULATIONS.doc

derreiter quasi random sequence generator, which has performed well in our earlier

Page 62: genetic_POPULATIONS.doc

tests [23], simple sequential inhibition (SSI) process [9] and nonaligned systematic

Page 63: genetic_POPULATIONS.doc

sampling, which originates from sampling design [34]. Collectively, we call different ways to generate initial populations point generators. The SSI process and the non- aligned systematic sampling are commonly used in statistics where point generators are called spatial point processes [9]. Spatial point processes are used, for example, in the statistical analysis of biological phenomena when simulating a distribution of a

Page 64: genetic_POPULATIONS.doc

population of plants or animals (see, e.g., [9, 19] and references therein).

Page 65: genetic_POPULATIONS.doc

Innumerical tests, we fix the geneticalgorithm and its parameter values and change

Page 66: genetic_POPULATIONS.doc

only initial populations.We use a test suite of 52 test functions from 16 different func- tion families and test whether the differences in best final objective function values

Page 67: genetic_POPULATIONS.doc

found are statistically significant between the different variants of genetic algorithms. We also study the convergence during the first generations by stopping the algorithm

Page 68: genetic_POPULATIONS.doc

prematurely after 10 and 20 generations.

Page 69: genetic_POPULATIONS.doc

The rest of the paper is organized as follows. In Section 2, we show that the initial

Page 70: genetic_POPULATIONS.doc

population may have an effect on the convergence of a genetic algorithm and give some further motivation for this work. We also shortly present the genetic algorithm used. In Section 3, we give an overview to different point generators that can be used

Page 71: genetic_POPULATIONS.doc

when generating initial populations, and in Section 4, we discuss and evaluate their speed and usability as well as the coverage and the genetic diversity of the points gen- erated. The numerical results of applying the different variants of genetic algorithms to our test suite are presented and analyzed in Section 5. In Section 6, we discuss the

Page 72: genetic_POPULATIONS.doc

results, and finally, in Section 7, we conclude the paper and present some directions

Page 73: genetic_POPULATIONS.doc

for future research.

Page 74: genetic_POPULATIONS.doc

2 Preliminaries

Page 75: genetic_POPULATIONS.doc

Population-based genetic algorithms (see, e.g., [15] and references therein) are de-

Page 76: genetic_POPULATIONS.doc

signed for solving problems that may have several local minima. They are very gen- eral problem solvers, which means that they can be used for solving a wide range of problems. On the other hand, they do not exploit problem-specific information,

Page 77: genetic_POPULATIONS.doc

which makes them less efficient. Hence, genetic algorithms ought to be used when

Page 78: genetic_POPULATIONS.doc

problem-specific methods are not available or if a wide range of problems need to be solved with a single algorithm. We consider global optimization problems of the

Page 79: genetic_POPULATIONS.doc

following form

Page 80: genetic_POPULATIONS.doc

minimize f (x)

Page 81: genetic_POPULATIONS.doc

subject to x l ≤ x ≤ xu i = 1, ... , n,

Page 82: genetic_POPULATIONS.doc

where f : R

n → R is the objective function, x are the decision variables and x , x

Page 83: genetic_POPULATIONS.doc

are the vectors containing the lower and upper bounds for the decision variables,

Page 84: genetic_POPULATIONS.doc

respectively.

Page 85: genetic_POPULATIONS.doc

10

J Glob Optim (2007) 37:405–436

(x /4000 )− (cos(x ) /

408

i=1

2 i

10

i=1

i √

i) i ≤ 100,

i=1

10

1 +i 30 | 2

k=1

k i

k i

2 k , i < 1.

Page 86: genetic_POPULATIONS.doc

A simple genetic algorithm includes three basic genetic operations: selection,

Page 87: genetic_POPULATIONS.doc

crossover and mutation. In selection, some solutions from the population are selected

Page 88: genetic_POPULATIONS.doc

as parents, in crossover the parents are crossbred to produce offspring and in muta- tion the offspring may be altered according to mutation rules. In genetic algorithms, solutions x are called individuals and iterations of an algorithm are called generations.

Page 89: genetic_POPULATIONS.doc

Many genetic algorithms also employ elitism, which means that a number of the best individuals are copied to the next population. We use a real-coded genetic algorithm that employs tournament selection, heuristic crossover, Michalewicz’s nonuniform

Page 90: genetic_POPULATIONS.doc

mutation and elitism. For further details of the genetic operations, see [25].

Page 91: genetic_POPULATIONS.doc

The algorithm used has the following parameters. Population size is the number

Page 92: genetic_POPULATIONS.doc

of individuals in a population and elitism size is the number of fittest individuals that are copied directly to the next generation. The fitness is evaluated using the objective function (fitness function) value f (x). Tournament size is the number of individuals randomly picked from the whole population for the tournament selection. Crossover rate and mutation rate are probabilities on which the parents are crossbred and off- spring mutated, respectively.Max generations, steps,andtolerance are parameters for the stopping criteria. The algorithm is stopped if a maximum number of generations (max generations) is reached or if there is no change (within the tolerance) in the best

Page 93: genetic_POPULATIONS.doc

objective function value during the last steps generations.

Page 94: genetic_POPULATIONS.doc

Genetic algorithms are Markov chains (see, e.g., [31, 46] and references therein).

Page 95: genetic_POPULATIONS.doc

Such chains are expected to converge to an equilibrium distribution independent of

Page 96: genetic_POPULATIONS.doc

the initial state. However, in many applications the state space of the chain (possible

Page 97: genetic_POPULATIONS.doc

point configurations in the feasible region) is extremely large and convergence can be very slow. This poses the question whether the number of the generations used is

Page 98: genetic_POPULATIONS.doc

sufficient in order to achieve the equilibrium. The initial configuration is one factor in

Page 99: genetic_POPULATIONS.doc

the speed of convergence. This is our motivation for studying empirically the role of

Page 100: genetic_POPULATIONS.doc

initial populations.

Page 101: genetic_POPULATIONS.doc

The difference in the early generationsbecomes important in solvingmany real-life

Page 102: genetic_POPULATIONS.doc

problems, where the evaluation of the objective functionmay requiretime-consuming simulations and the optimization algorithm may have to be stopped prematurely. We

Page 103: genetic_POPULATIONS.doc

also expect that the influence of the initial population may carry further, in terms of generations, when solving of the problem requires a large number of function

Page 104: genetic_POPULATIONS.doc

evaluations.

Page 105: genetic_POPULATIONS.doc

Figures 1 and 2 illustrate the convergence of a genetic algorithm for the 10-dimen-

Page 106: genetic_POPULATIONS.doc

sional Griewangk function and 10-dimensional Katsuura function:

Page 107: genetic_POPULATIONS.doc

Griewangk function : f (x) =

−10 ≤ x

Katsuura function: f (x) =

x − 2 x |−0.1 < x

Page 108: genetic_POPULATIONS.doc

Each curve illustrates the average convergence for 100 separate runs of a genetic algorithm. The three different curves in Figs. 1 and 2 correspond to runs with differ- ent initial populations generated by using a pseudo random number generator. Here pseudo stands for the case, where the initial pseudo random population is spread out over the whole feasible region. Furthermore, clustered 1 and clustered 2 stand for

Page 109: genetic_POPULATIONS.doc

cases where the initial population is restricted to a subspace of the feasible region. The

Page 110: genetic_POPULATIONS.doc

1500

1000

500

J Glob Optim (2007) 37:405–436 409

5 10 15 20 0

50

60

Bes

t obj

ectiv

e fu

nctio

n va

lue

foun

d

10

20

30

40

clustered 1 pseudo clustered 2

45 50 55 60 40 0

2000

Bes

t obj

ectiv

e fu

nctio

n va

lue

foun

d

pseudo

i

Page 111: genetic_POPULATIONS.doc

Generations

Page 112: genetic_POPULATIONS.doc

Fig. 1 The convergence of a genetic algorithm for 10D Griewangk function with different initial

Page 113: genetic_POPULATIONS.doc

populations

Page 114: genetic_POPULATIONS.doc

clustered 1

clustered 2

Page 115: genetic_POPULATIONS.doc

Generations

Page 116: genetic_POPULATIONS.doc

Fig. 2 The convergence of a genetic algorithm for 10D Katsuura function with different initial

Page 117: genetic_POPULATIONS.doc

populations

Page 118: genetic_POPULATIONS.doc

subspace is defined by restricting each variable x to the upper and lower80% of their total range for clustered 1 and clustered 2, respectively. For theGriewangk function the

Page 119: genetic_POPULATIONS.doc

curves merge after 30 generations. The Katsuura function requires more generations

Page 120: genetic_POPULATIONS.doc

and, therefore, also the curves merge later than for the Griewangk function. (Note

Page 121: genetic_POPULATIONS.doc

the different scales in Figs. 1 and 2.) For both Griewangk and Katsuura functions, the differences in the best objective function values found is quite significant when the number of generations is small. This indicates that the initial population has an effect

Page 122: genetic_POPULATIONS.doc

on the convergence of a genetic algorithm.

Page 123: genetic_POPULATIONS.doc

The clustered populations used in the simple example above were only theoretical.

Page 124: genetic_POPULATIONS.doc

In practice, such initial populations would hardly be used, unless there were some

Page 125: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

Fig. 3 Coverage of the connecting lines with different patterns

410

Page 126: genetic_POPULATIONS.doc

a priori knowledge about the location of the global minima. The rest of the paper

Page 127: genetic_POPULATIONS.doc

concentrates on more realistic alternative ways for generating initial populations.

Page 128: genetic_POPULATIONS.doc

When there is no a priori information available on the location and the number

Page 129: genetic_POPULATIONS.doc

of local optima, then the initial population of a genetic algorithm should be able to reach as large part of the feasible region as possible by means of crossover. We call this property genetic diversity, and it is related to the independence of points.Another desirable property for an initial population is a good uniform coverage. By a good uniform coverage we mean that the points are well spread out to cover the whole fea- sible region. Points have a gooduniform coverage if they do notform clusters or leave relatively large areas of the feasible region unexplored. A good uniform coverage is desired, because then information is obtained throughout the whole feasible region.

Page 130: genetic_POPULATIONS.doc

This helps to prevent premature convergence.

Page 131: genetic_POPULATIONS.doc

Genetic diversity and a gooduniform coverage are in practice conflicting objectives,

Page 132: genetic_POPULATIONS.doc

because an optimal uniform coverage is often achieved using systematic structures,

Page 133: genetic_POPULATIONS.doc

which limit the set of possible offspring. For example, it is known that for a closely related problem of sphere packing a triangular grid provides the optimal packing in

Page 134: genetic_POPULATIONS.doc

a 2-dimensional case (see, e.g., [44]). Hence, if the pattern was repeated, the distance between points would reach its maximum, when using triangular grid. However, the set of possible offspring is limited as illustrated in Fig. 3. This happens also if the

Page 135: genetic_POPULATIONS.doc

sample points form clusters. Therefore, seemingly random (independent) point sets with no clustering are preferred, since they provide information over the whole fea- sible region and a large part of the feasible region can be reached by the means of crossover. Figure 3 illustrates a rectangular grid, triangular grid, clustered points, and seemingly random points with no clustering. In the figure, the points marked with

Page 136: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 411

ences.

paper.

Page 137: genetic_POPULATIONS.doc

small black dots are parent solutions and the lines running through them illustrate all

Page 138: genetic_POPULATIONS.doc

the possible locations of offsprings.

Page 139: genetic_POPULATIONS.doc

3 Generating initial population

Page 140: genetic_POPULATIONS.doc

Many sub-areas of geneticalgorithms have been studied elaborately, but the selection

Page 141: genetic_POPULATIONS.doc

of an initial population has been widely ignored.Asmentioned earlier, the traditional

Page 142: genetic_POPULATIONS.doc

way to generate the initial population is to use pseudo random numbers, and more recently also quasi random sequences have been applied in [23, 24]. Pseudo random numbers imitate genuine random numbers and quasi random sequences are designed

Page 143: genetic_POPULATIONS.doc

to produce points that maximally avoid each other.

Page 144: genetic_POPULATIONS.doc

Pseudo random initial populations can be generated in numerous ways. The main

Page 145: genetic_POPULATIONS.doc

classes are congruential and recursive generators. Common congruential generators

Page 146: genetic_POPULATIONS.doc

include linear, quadratic, inversive, additive and parallel linear congruential genera- tors [14, 45]. Recursive generators include multiplicative recursive, lagged Fibonacci,

Page 147: genetic_POPULATIONS.doc

multiply-with-carry-generator, add-with-carry and substract-with-borrow generators [14]. There are also pseudo random vector generators, which produce sequences of

Page 148: genetic_POPULATIONS.doc

vectors instead of scalars.Examples of those are feedback shift register generator [14]

Page 149: genetic_POPULATIONS.doc

and SQRT generator [43].

Page 150: genetic_POPULATIONS.doc

Common quasi random sequence generators include Van der Corput, Hammers-

Page 151: genetic_POPULATIONS.doc

ley, Halton, Faure, Sobol’ and Niederreiter generators [4, 14, 30]. For the convenience of the reader some examples of both pseudo random number generators and quasi

Page 152: genetic_POPULATIONS.doc

random sequence generators are included in the Appendix along with further refer-

Page 153: genetic_POPULATIONS.doc

Pseudo random numbers and quasi random sequences are quite well-known for

Page 154: genetic_POPULATIONS.doc

researchers in optimization, but there are also other types of point generators. Spatial point processes are commonly used in statistics but they are less well-known in opti-

Page 155: genetic_POPULATIONS.doc

mization. Therefore, we now shortly describe the spatial point processes used in this

Page 156: genetic_POPULATIONS.doc

3.1 Spatial point processes

Page 157: genetic_POPULATIONS.doc

Spatial point processes [9] can be considered to betransformations of pseudorandom point initialization which lead to a good uniform coverage and simultaneously avoid

Page 158: genetic_POPULATIONS.doc

periodicity in the configuration generated.

Page 159: genetic_POPULATIONS.doc

Some spatial point processes include a parameter, by which the proportion of the

Page 160: genetic_POPULATIONS.doc

two conflicting properties, genetic diversity and a good uniform coverage, can be con- trolled. Hence, the same process can be used to generate genetically diverse points or

Page 161: genetic_POPULATIONS.doc

points with a good uniform coverage or something in between.

Page 162: genetic_POPULATIONS.doc

There are several types of spatial point processes. For our purposes, it is ade-

Page 163: genetic_POPULATIONS.doc

quate to differentiate between clustering and inhibition processes. Clustering pro- cesses generate points that simulate populations, where the individuals tend to form clusters. However, we are more interested in the situation where each individual is as far apart as possible from the adjacent individuals resulting in an evenly dis- tributed population without clusters. Therefore, we concentrate on inhibition pro-

Page 164: genetic_POPULATIONS.doc

cesses, where an individual (a point) is either explicitly prohibited to be located

Page 165: genetic_POPULATIONS.doc

closer than some predefined minimum distance > 0 to the other individuals, or individuals are kept apart by some implicit means. We describe here two inhibition

Page 166: genetic_POPULATIONS.doc

S =

J Glob Optim (2007) 37:405–436

j

n k =1 k

= 0, ... , bn −1.

l = (S)mod b

412

End do

n

i,1 j,2 ] ) ,

1. Do i i

Do j i

v −v j

x l, j j )End do

Page 167: genetic_POPULATIONS.doc

processes: a simple sequential inhibition process and a nonaligned systematic

Page 168: genetic_POPULATIONS.doc

sampling.

Page 169: genetic_POPULATIONS.doc

Simple sequential inhibition process: In the simple sequential inhibition (SSI)

Page 170: genetic_POPULATIONS.doc

process [9], a new individual is accepted to enter the population only if its distance to

Page 171: genetic_POPULATIONS.doc

all the existing individuals in the population is at least .

Page 172: genetic_POPULATIONS.doc

In the following pseudo code for the SSI process the parameter population size is

Page 173: genetic_POPULATIONS.doc

the number of points to be generated, futile is the number of rejected trial points dur- ing the current iteration, Max_futile is the maximum number of rejected trial points in one iteration before terminating the process, and n is the dimension of the space

Page 174: genetic_POPULATIONS.doc

where the points are generated.

Page 175: genetic_POPULATIONS.doc

0. Initialize parameters population size,Max_futile and dimension n and set futile = 0

Page 176: genetic_POPULATIONS.doc

and k = 0.

Page 177: genetic_POPULATIONS.doc

1. Do until k = = population size or futile==Max_futile

Page 178: genetic_POPULATIONS.doc

1.1 Generate a trial point.

Page 179: genetic_POPULATIONS.doc

(Use a pseudo or a quasi random number generator to generate a trial point

Page 180: genetic_POPULATIONS.doc

to the n-dimensional unit hyper cube.)

Page 181: genetic_POPULATIONS.doc

1.2 Check whether the trial point is accepted.

Page 182: genetic_POPULATIONS.doc

If the distance to the existing individuals is larger than , accept the point

Page 183: genetic_POPULATIONS.doc

into the population, and set k = k + 1 a nd futile= 0.

Page 184: genetic_POPULATIONS.doc

Else, set futile =futile+1.

Page 185: genetic_POPULATIONS.doc

The distance between points can becomputed using differentmetrics. If theminimum

Page 186: genetic_POPULATIONS.doc

distance is defined to be too large, then themaximumnumber of rejected trial points in one iteration is reached and the process is terminated prematurely. In that case, we

Page 187: genetic_POPULATIONS.doc

supplement the sample with pseudo random points to match the population size.

Page 188: genetic_POPULATIONS.doc

Nonaligned systematic sampling: In the nonaligned systematic sampling, which orig- inates from sampling design [34], the unit hyper cube is divided into b elementary

Page 189: genetic_POPULATIONS.doc

intervals (see Appendix) with equal side lengths (i.e., an equally spaced grid). Then one sample point is selected from each elementary interval according to prescribed

Page 190: genetic_POPULATIONS.doc

rules. Nonaligned systematic sampling uses b · n pseudo random numbers to define

Page 191: genetic_POPULATIONS.doc

the location of the sample points. In two dimensions, the sample points are generated

Page 192: genetic_POPULATIONS.doc

using the formula

Page 193: genetic_POPULATIONS.doc

x = ([(j −1)+r ] , [( i −1)+r

Page 194: genetic_POPULATIONS.doc

where i, j = 1, ... , b, = 1 / b , a nd r is a b ×n array of pseudo random numbers, see

Page 195: genetic_POPULATIONS.doc

Fig. 4. In the following pseudo code for an n-dimensional case of the nonaligned sys- tematic sampling, v is a vector of auxiliary integer variables. It is used to determine the correct elementary interval and the correct pseudo random number when calculating

Page 196: genetic_POPULATIONS.doc

the nonaligned systematic sample points x.

Page 197: genetic_POPULATIONS.doc

0. Initialize v = 0forj = 0, ... , n −1, and the pseudo random number array .

Page 198: genetic_POPULATIONS.doc

1.1 Compute sample point x

Page 199: genetic_POPULATIONS.doc

= 0, ... , n −1 // Compute the jth component of x

Page 200: genetic_POPULATIONS.doc

i

j = (r +v

Page 201: genetic_POPULATIONS.doc

Fig. 4 Nonaligned systematic sampling points in two dimensions

1.2 Update v Do j = 0, ... , n −1

j = (v +1)mod b

If ((i)mod b −1) then v

413

j = = bj j

End do End do

2

Page 202: genetic_POPULATIONS.doc

Note that the number of nonaligned systematic sampling points cannot be chosen

Page 203: genetic_POPULATIONS.doc

freely, but is determined by the grid size and the dimension. In practice, when a cer-

Page 204: genetic_POPULATIONS.doc

tain number of sample points is required, we supplement the nonaligned systematic

Page 205: genetic_POPULATIONS.doc

sample with pseudo random points to match the required number of points.

Page 206: genetic_POPULATIONS.doc

4 Properties of point generators

Page 207: genetic_POPULATIONS.doc

In this section, we consider the distribution and the genetic diversity of the points

Page 208: genetic_POPULATIONS.doc

generated by pseudo random number generators, quasi random sequence generators

Page 209: genetic_POPULATIONS.doc

and spatial point processes. We also consider the speed and the usability of some specific generators. We are interested in point sets with a relatively small number of points since the number of points generated equals the population size and ranges

Page 210: genetic_POPULATIONS.doc

typically from tens to a few hundreds.

Page 211: genetic_POPULATIONS.doc

In the illustrations and numerical examples of point generation, we have selected

Page 212: genetic_POPULATIONS.doc

good representatives for the different types of point generators. The pseudo random

Page 213: genetic_POPULATIONS.doc

number generator is a multiplicative linear congruential generator from the well-

Page 214: genetic_POPULATIONS.doc

established numerical library of the Numerical Algorithms Group Ltd (NAG) and the quasi random sequence generator is the Niederreiter generator [30] that has

Page 215: genetic_POPULATIONS.doc

proved successful in our earlier tests [23, 24]. The representatives of spatial point

Page 216: genetic_POPULATIONS.doc

processes are the SSI process and the nonaligned systematic sampling. In the SSI pro-

Page 217: genetic_POPULATIONS.doc

cess, the trial points are generated using the Niederreiter generator and the distances

Page 218: genetic_POPULATIONS.doc

are measured using L -metric. The nonaligned systematic sampling is as defined in

Page 219: genetic_POPULATIONS.doc

Section 3.

Page 220: genetic_POPULATIONS.doc

The properties of the point generators are described in the following four sub-

Page 221: genetic_POPULATIONS.doc

sections and the results are summarized and some conclusions are drawn in the fifth

Page 222: genetic_POPULATIONS.doc

subsection.

Page 223: genetic_POPULATIONS.doc

as the number of points x ,1

J Glob Optim (2007) 37:405–436 414

n ⊂ Rn 1 N ∈In

n

k N

n N B

A(B ;P)N

n −1

m

Page 224: genetic_POPULATIONS.doc

4.1 Distribution

Page 225: genetic_POPULATIONS.doc

Intuitively thinking, it is beneficial if no large areas are left unexplored when sampling individuals for an initial population of a genetic algorithm. The point generators pre-

Page 226: genetic_POPULATIONS.doc

sented in Section 3 produce point clouds with a degree ofuniform coverage.However, there are differences in how well the points of these sequences are spread out. In opti- mization and in numerical integration, the goodness of the uniform coverage of a

Page 227: genetic_POPULATIONS.doc

point set is commonly measured by discrepancy or dispersion [30]. Here, we give the

Page 228: genetic_POPULATIONS.doc

definition for the discrepancy.

Page 229: genetic_POPULATIONS.doc

Discrepancy Let I be an n-dimensional unit hyper cube, let P be a set consist-

ing of points x , ..., x ,andletB be a nonempty family of Lebesgue-measurable

Page 230: genetic_POPULATIONS.doc

subintervals of I and B ∈B. Furthermore, let A(B;P)be a counting function defined

Page 231: genetic_POPULATIONS.doc

≤ k ≤ N, for which xk ∈ B. Then, discrepancy D

Page 232: genetic_POPULATIONS.doc

with respect to P in I is defined as D (B;P) = sup ∈B − λ ( B ) , w here λ is a

Page 233: genetic_POPULATIONS.doc

Lebesgue-measure.

Page 234: genetic_POPULATIONS.doc

Discrepancy is large, when there exists clusters or large unexplored areas. Hence,

Page 235: genetic_POPULATIONS.doc

we are interested in point sets that have low values for discrepancy. A mathematical relationship between discrepancy and dispersion is given in [30], and it shows that

Page 236: genetic_POPULATIONS.doc

every low-discrepancy sequence is also a low-dispersion sequence, but not vice versa.

Page 237: genetic_POPULATIONS.doc

Quasi random sequences are also called low-discrepancy sequences. Their discrep-

Page 238: genetic_POPULATIONS.doc

ancy value for a large sample size N is of order of magnitude C(logN) N , w here

Page 239: genetic_POPULATIONS.doc

C is a generator-specific coefficient depending only on the dimension n [27]. This is also a minimum possible discrepancy size for large N (see, e.g., [27]). The bounds for discrepancy, however, are relevant only for a very large number of sample points

Page 240: genetic_POPULATIONS.doc

as used in numerical integration. When generating only initial populations, we are more interested in the distribution of a small number of points. Some quasi random

Page 241: genetic_POPULATIONS.doc

sequences have the property that the first b successive points divide evenly on the correspondingelementary intervals (seeAppendix). This indicates that quasirandom sequencesmay have good discrepancy values also forsmall sample sizes. Figure 5 illus-

Page 242: genetic_POPULATIONS.doc

trates four two-dimensional initial populations with 1,024 individuals generated using the multiplicative linear congruential (pseudo) generator, the Niederreiter generator,

Page 243: genetic_POPULATIONS.doc

the SSI process and the nonaligned systematic sampling, respectively.

Page 244: genetic_POPULATIONS.doc

In Fig. 5, we see that pseudo random points form clusters and leave some areas

Page 245: genetic_POPULATIONS.doc

relatively unexplored, whereas the other samples cover the feasible region quite well.

Page 246: genetic_POPULATIONS.doc

However, quasi random sequences may sometimes form point patterns, whose dis- tribution properties depend on the number of points generated. Then, at a certain number of sample points the pattern may be unfavorable for optimization purposes. Anexample of this behavior is illustrated in Fig. 6, where again 1,024 points were gen- eratedwith a slightly different configuration. Formore information on the distribution

Page 247: genetic_POPULATIONS.doc

of different quasi random sequences, we refer to [27].

Page 248: genetic_POPULATIONS.doc

Next, we consider the distribution of small sample sizes. In optimization with

Page 249: genetic_POPULATIONS.doc

geneticalgorithms we used a population of 201 individuals. Therefore, we nextempir- ically examine the distribution properties of sets with 201 points generated in four different ways. Henceforth, in tables and figures, we will use abbreviations Pseudo,

Page 250: genetic_POPULATIONS.doc

Nieder, SSI and Nonalig for the representatives selected from different types of point

Page 251: genetic_POPULATIONS.doc

generators (see the beginning of this section).

Page 252: genetic_POPULATIONS.doc

415 J Glob Optim (2007) 37:405–436

Pseudo generator Niederreiter generator

0.5 1 0 0

0.5

1

0.5 1 0 0

1

0.5

0.5 1 0 0

0.5

1

0.5 1 0 0

0.5

1

Page 253: genetic_POPULATIONS.doc

SSI process

Nonaligned systematic sampling

Page 254: genetic_POPULATIONS.doc

Fig. 5 Point sets of 1,024 points with different generators

Page 255: genetic_POPULATIONS.doc

We consider 2 and 10-dimensional cases to show how the distribution properties

Page 256: genetic_POPULATIONS.doc

may change with the dimension. First, we use an empirical empty space statistic to examine how well the sample points cover the feasible region [34]. Empty space

Page 257: genetic_POPULATIONS.doc

statistic function ess is defined as

Page 258: genetic_POPULATIONS.doc

ess(r) = 1 −Pr(B(x, r) is empty),

Page 259: genetic_POPULATIONS.doc

where x is a randomly chosen point, B(x, r) is an x-centered ball with radius r and Pr denotes the probability. Hence, assume there is a point set S and further assume there is a ball B with radius r and whose center point is randomly chosen from the feasible region (in our case, the unit hyper cube). Then, for each radius r, the empty space statistic function tells what the probability is that the ball B is empty, that is, the

Page 260: genetic_POPULATIONS.doc

intersection of the ball B and the point set S is empty.

Page 261: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

Fig. 6 Another example of 1,024 points generated with the Niederreiter generator

sets.

1

0.5

0

Niederreiter generator

0.5

1

1 To experimentally define the empty space statistic functions for each point set con-

0

1

results.

Page 262: genetic_POPULATIONS.doc

taining 201 points, we generated 10,000 auxiliary pseudo random points on a unit hyper cube and for each of the 10,000 auxiliary points calculated the maximal radius

Page 263: genetic_POPULATIONS.doc

r so that B(x, r) is empty, that is, does not contain any of the 201 points under consid- eration. We did this in both 2- and 10-dimensional cases for the four different point

Page 264: genetic_POPULATIONS.doc

The empirical empty space statistic functions are illustrated in Fig. 7. For us, the

Page 265: genetic_POPULATIONS.doc

important property of the empty space statistic function is the steepness of the curve

Page 266: genetic_POPULATIONS.doc

(the steeper the better). If the curve is steep, it indicates that if a random ball with radius r is selected from the unit hyper cube, then whether the ball is empty depends mainly on the radius—not the location of the ball. If, on the other hand, the curve is gentle, the emptiness of the ball depends strongly on the location, hence the point set

Page 267: genetic_POPULATIONS.doc

is not evenly distributed.

Page 268: genetic_POPULATIONS.doc

The empiricalempty space statistic functions show that in two dimensions, the non-

Page 269: genetic_POPULATIONS.doc

aligned systematic sampling and the simple sequential inhibition (SSI) process provide

Page 270: genetic_POPULATIONS.doc

the best coverages closely followed by the Niederreiter quasi random sequence gen- erator. The pseudo random initial population has clearly the worst coverage, which

Page 271: genetic_POPULATIONS.doc

confirms the findings earlier shown in Fig. 5. However, in ten dimensions the differ-

Page 272: genetic_POPULATIONS.doc

ences diminish, because the populations become sparse. The only notable difference is that for the population generated with the SSI process the curve of the empirical

Page 273: genetic_POPULATIONS.doc

empty space statistic function is now below the other curves with the small values of

Page 274: genetic_POPULATIONS.doc

r but rises more steeply in the end. This indicates better coverage for the SSI process.

Page 275: genetic_POPULATIONS.doc

4.2 Genetic diversity

Page 276: genetic_POPULATIONS.doc

The population of a genetic algorithm evolves largely by crossovers and mutations. In our implementation, the most dominant genetic operator is crossover, since it usually changes the solutions most. We define genetic diversity as a property by which the

Page 277: genetic_POPULATIONS.doc

Here, we naturally used a different pseudo random number generator than when generating the

Page 278: genetic_POPULATIONS.doc

initial population. To make sure that the use of pseudo random numbers did not bias the results we also computed the empty space statistic functions using a grid of 10,000 points, which led to similar

Page 279: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 417

0.05 0.1 0.15 0.2

ess

(r)

2D

0 0

0.2

0.4

0.6

0.8

1

Pseudo Nieder SSI Nonalig

0.2 0.4 0.6 0.8 1 1.2 1.4 0 0

0.2

0.4

0.6

0.8

1

ess

(r)

10D

Pseudo Nieder SSI Nonalig

n 1 n ,

n n

Page 280: genetic_POPULATIONS.doc

Radius r

Radius r

Page 281: genetic_POPULATIONS.doc

Fig. 7 Empirical empty space statistic functions

Page 282: genetic_POPULATIONS.doc

genetic algorithm is able to reach as large a part of the feasible region as possible by

Page 283: genetic_POPULATIONS.doc

means of crossover. Genetically more diverse initial populations are preferable.

Page 284: genetic_POPULATIONS.doc

Next,we concentrate on the independence of points,which is onemain criterion for

Page 285: genetic_POPULATIONS.doc

genetic diversity.Asmentioned earlier, truly independent points cannot be generated

Page 286: genetic_POPULATIONS.doc

algorithmically. The pseudorandomnumbers try to imitate independentnumbers, yet they are deterministically generated by an algorithm. For example, it has been shown that the pseudo random points generated by a linear congruential generator lie on a simple lattice (see, e.g., [14] and references therein). It is also known that bad choices for the coefficients of a linear congruential generator may result in a very bad lattice structure (see, e.g., [14]). The choices for the coefficients play a crucial role for all the

Page 287: genetic_POPULATIONS.doc

pseudo random number generators.

Page 288: genetic_POPULATIONS.doc

Also quasi random sequences are generated algorithmically. However, the idea of

Page 289: genetic_POPULATIONS.doc

quasi random points is not to imitate random points but to try to cover the whole

Page 290: genetic_POPULATIONS.doc

search space yet maintaining a certain degree of randomness. These initial settings speak for a better genetic diversity for pseudo random initial populations. In empiri-

Page 291: genetic_POPULATIONS.doc

cal tests, when comparing pseudo and quasi initial populations, we can study plots of 2-dimensional populations or plots of n-dimensional populations that are projected

Page 292: genetic_POPULATIONS.doc

to two dimensions. In Figs. 5 and 6 we can notice that the Niederreiter generator

Page 293: genetic_POPULATIONS.doc

produces points that have a more distinguishable pattern whereas pseudo random

Page 294: genetic_POPULATIONS.doc

points have a pattern that seems more random.

Page 295: genetic_POPULATIONS.doc

We studied the independence of point sets of 201 2-dimensional points using Rip-

Page 296: genetic_POPULATIONS.doc

ley’s K-function and L-function [34]. The K-function is defined as follows: λK(r) is the expected number of further points within a ball with radius r and with a center at a randomly chosen point ( λ is the expected number of points in a unit hyper cube).

Page 297: genetic_POPULATIONS.doc

L-function is obtained from the K-function through transformation

Page 298: genetic_POPULATIONS.doc

L(r) = (K(r) / )

Page 299: genetic_POPULATIONS.doc

where is the volume of the unit ball in R . Figure 8 illustrates the L-functions for the pseudo random points, the Niederreiter quasi random points, points from the SSI process and the nonaligned systematic sampling points. For independent points

Page 300: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

Four L−functions

418

0.05 0.1 0.15 0.2 0.25 0 0

0.05

0.1

0.15

0.2

0.25

Distance r

L fu

nctio

n

Pseudo Nieder SSI Nonalig

we get 2

Page 301: genetic_POPULATIONS.doc

Fig. 8 L-functions for

Page 302: genetic_POPULATIONS.doc

different initial populations

Page 303: genetic_POPULATIONS.doc

the L-function is a straight line L(r)=r. We can see that pseudo random points imi- tate well independent points, whereas Niederreiter quasi random points and the two

Page 304: genetic_POPULATIONS.doc

spatial point processes lack the points falling close to each other. Based on these

Page 305: genetic_POPULATIONS.doc

observations, we conclude that the pseudo random generator produces genetically

Page 306: genetic_POPULATIONS.doc

more diverse points than the other point generators.

Page 307: genetic_POPULATIONS.doc

The points generated by the SSI process are essentially pseudo random points.

Page 308: genetic_POPULATIONS.doc

However, the trial points falling too close to each other are not accepted to the popu- lation and this affects the genetic diversity. The genetic diversity versus a gooduniform coverage can be controlled using theminimum distanceparameter . In what follows, we use the term probabilistic maximal minimum distance (PMMD) and we should find the value of PMMD so that a desired number of points can fit in the feasible region with a given probability. In the 2-dimensional example in Fig. 8, we can see

Page 309: genetic_POPULATIONS.doc

that, except for small distances, the points of the SSI process fall on a straight line.

Page 310: genetic_POPULATIONS.doc

The nonaligned systematic sampling uses only n ·b pseudo random points, where

Page 311: genetic_POPULATIONS.doc

n is the dimension and 1/b is the side length of the elementary interval (the distance

Page 312: genetic_POPULATIONS.doc

between two adjacent grid points). Therefore, there is not much genetic diversity in nonaligned systematic sample points. In Fig. 8, we see this as an oscillating behavior of the L-function. Furthermore, the number of nonaligned systematic sample points increases quickly with the dimension if the grid size is kept constant. For example, in

Page 313: genetic_POPULATIONS.doc

10 or 20-dimensional cases, even if we use only two grid points along each dimension,

Page 314: genetic_POPULATIONS.doc

10 = 1, 024 and 220 = 1, 048, 576 sample points, respectively. When using a

Page 315: genetic_POPULATIONS.doc

population size of 201 individuals, we get in practice only one genuinesystematicsam-

Page 316: genetic_POPULATIONS.doc

ple point in 8 or more dimensions (the rest are supplemented pseudo random points). This means that, in large dimensions, the nonaligned systematic sampling reduces to

Page 317: genetic_POPULATIONS.doc

pseudo random sampling.

Page 318: genetic_POPULATIONS.doc

4.3 Speed

Page 319: genetic_POPULATIONS.doc

For some applications the speed of the generator may be of importance. The CPU times for the tested generators and the different numbers of points generated in dimensions 2, 5, 10, 20, 50 are shown in Table 1. The test runs were performed on

Page 320: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 419

Dim Points SSI Nonalig

2 201 0.00 0.00 0.06 0.00 4 0.00 17.60 0.02 5 0.02 1774.0 0.23

5 201 0.00 0.00 0.2 0.00 4 0.00 17.84 0.06 5 0.03 1760.0 0.67

10 201 0.00 0.00 0.23 0.00 10 10 4 0.03 0.01 17.71 0.04 10 10 5 0.28 0.05 1773.0 0.98 20 201 0.00 0.01 0.73 0.00 20 10 4 0.06 0.01 10.78 0.05 20 10 5 0.55 0.09 21.33 0.46 50 201 0.01 0.00 1.54 0.01 50 10 4 0.14 0.02 44.50 0.12 50 10 5 1.38 0.2 45.68 1.2

and 50.

Page 321: genetic_POPULATIONS.doc

Tab le 1 CPU times in seconds

Pseudo Nieder

Page 322: genetic_POPULATIONS.doc

2 1 00.01

Page 323: genetic_POPULATIONS.doc

2 1 00.06

Page 324: genetic_POPULATIONS.doc

5 1 00.02

Page 325: genetic_POPULATIONS.doc

5 1 00.14

Page 326: genetic_POPULATIONS.doc

an HP9000/J5600 computer, and the CPU times were obtained using NAG routine X05BAF, which returns the CPU time with the accuracy of 0.01 s. For this reason, differences betweensome generators become evident only when using a largenumber

Page 327: genetic_POPULATIONS.doc

of points. The columnDim in Table 1 indicates thedimensions of the points generated

Page 328: genetic_POPULATIONS.doc

and the column Points lists the number of points generated. The abbreviations for the

Page 329: genetic_POPULATIONS.doc

generators are the same as earlier, and the CPU times are given in seconds.

Page 330: genetic_POPULATIONS.doc

In Table 1, we see that the SSI process is by far the slowest resulting from the

Page 331: genetic_POPULATIONS.doc

implementation which emphasizes the good coverage at the expense of speed. Our implementation allows a maximum of 2,000 futile trial points at each iteration before

Page 332: genetic_POPULATIONS.doc

terminating the process. The speed (and the distribution of points) for the SSI process depends strongly on the probabilistic maximal minimum distance PMMD. If PMMD

Page 333: genetic_POPULATIONS.doc

is selectedmaximally large (i.e., PMMD is as large as possible yet so that the algorithm does not have to supplement the population with pseudo random points, see Section 3.1), then a good coverage of the points is emphasized and the number of futile tries

Page 334: genetic_POPULATIONS.doc

increases slowing down the process. Our implementation of the SSI process auto-

Page 335: genetic_POPULATIONS.doc

matically estimates the maximally large PMMD. The values in Table 1 indicate that

Page 336: genetic_POPULATIONS.doc

the estimate for PMMD is probably not good (distribution-wise) for dimensions 20

Page 337: genetic_POPULATIONS.doc

Table 1 shows that the implementation of the Niederreiter quasi random number

Page 338: genetic_POPULATIONS.doc

generator is the fastest followed closely by the pseudo random number generator and the nonaligned systematic sampling process. In optimization, where only a relatively

Page 339: genetic_POPULATIONS.doc

small number of points is generated, there is no practical difference between the speed of these three generators. For pseudo and quasi random number generators

Page 340: genetic_POPULATIONS.doc

the speed seems to be directly proportional to the number of points generated. We can see that, compared to other generators, the speed of the nonaligned systematic

Page 341: genetic_POPULATIONS.doc

sampling grows for largerdimensions. This can be explained by different ratios of gen- uine nonaligned sampling points. As mentioned earlier, the population of nonaligned

Page 342: genetic_POPULATIONS.doc

systematic sampling points is supplemented by pseudo random points to match the

Page 343: genetic_POPULATIONS.doc

required population size. Hence, for example, in 5 dimensions there are 78% and

Page 344: genetic_POPULATIONS.doc

100% of genuine nonaligned systematic sampling points in sets of 10,000 and 100,000

Page 345: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 420

20 6

here.

Page 346: genetic_POPULATIONS.doc

sample points, respectively, whereas in 10 dimensions the respective values are only 10% and59%. In 20 dimensions, all except onesample point are already random even

Page 347: genetic_POPULATIONS.doc

if the sample size was as large as one million (2 >10 ).

Page 348: genetic_POPULATIONS.doc

4.4 Usability

Page 349: genetic_POPULATIONS.doc

In this section, we discuss the usability of different number generators. Some fea-

Page 350: genetic_POPULATIONS.doc

tures affecting usability have already been mentioned earlier and are summarized

Page 351: genetic_POPULATIONS.doc

Pseudo random number generators are by far the most commonly used ones in

Page 352: genetic_POPULATIONS.doc

optimization. The greatest advantage of pseudo random number generators concern- ing usability is that there exist implementations that are well-established, well-tested

Page 353: genetic_POPULATIONS.doc

and easily available.However, a goodrandom-like behavior is not amatter-of-course,

Page 354: genetic_POPULATIONS.doc

but requires a careful choice of parameters, that is, coefficients and seeds (or initial sequences). As earlier noted for linear congruential generators, wrong choices for parameter values may cause the points in a sequence to be badly distributed. This is true for other generators as well. Many of the pseudo random number generators are easy to implement, but the analysis of the distribution is difficult, especially for an arbitrary seed (or arbitrary initial sequences). According to [14], only well-tested pseudo random generators should be used. These generators often use fixed coeffi- cients and sometimeslimit the choice of seeds (or initial sequences). This is one way to

Page 355: genetic_POPULATIONS.doc

secure good distribution properties, assuming those values are tested and approved.

Page 356: genetic_POPULATIONS.doc

Contrary to pseudo random number generators, quasi random number generators

Page 357: genetic_POPULATIONS.doc

are neither so well-established, well-tested nor easily available. On the other hand, quasi random number generators having solid theoretical properties do not need so

Page 358: genetic_POPULATIONS.doc

much numerical testing. The use of available quasi generators is easy and the quasi

Page 359: genetic_POPULATIONS.doc

random points are guaranteed to have some advantageous properties as described in the Appendix. Quasi generators provide the same sequences on different runs of the

Page 360: genetic_POPULATIONS.doc

generator and, hence, they do not require a seed or an initial sequence from the user.

Page 361: genetic_POPULATIONS.doc

Since quasi generators take into consideration the location of the previous points, they also expect that the sequence is startedfrom the beginning. These two properties make quasi generators easier to use, because the user only has to give the number of points to be generated and the dimension as parameters. At the same time, these properties may become a disadvantage if the user for some reason wants to obtain

Page 362: genetic_POPULATIONS.doc

different sequences in the same dimension. On this occasion, the easiest way out might be to generate points in higher dimensions and then projectthem to the desired

Page 363: genetic_POPULATIONS.doc

dimension, but this may have an effect on the distribution properties.

Page 364: genetic_POPULATIONS.doc

The SSI process is not commonly available for an n-dimensional case, but the

Page 365: genetic_POPULATIONS.doc

implementation is straightforward given that the user provides the PMMD value.

Page 366: genetic_POPULATIONS.doc

However, it gets more complicated if the user provides only the number of points to

Page 367: genetic_POPULATIONS.doc

be generated and theoptimalPMMD(with respect to the coverage)must beestimated

Page 368: genetic_POPULATIONS.doc

automatically. This issue is discussed in more detail in Section 6. An interesting prop- erty of the SSI process is that the proportion of the two conflicting objectives, that is,

Page 369: genetic_POPULATIONS.doc

the genetic diversity and the good uniform coverage, can both be controlled using the

Page 370: genetic_POPULATIONS.doc

probabilistic maximal minimum distance parameter PMMD. Note that, if desired, the SSI process provides different points on every run just like pseudo random number

Page 371: genetic_POPULATIONS.doc

generators.

Page 372: genetic_POPULATIONS.doc

The nonaligned systematic sampling processes are not commonly available in

Page 373: genetic_POPULATIONS.doc

n-dimensions, but they are easy to implement and to use. However, when

Page 374: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 421

Coverage + ++ +++ +++/+ +++ ++ ++ +

Speed +++ +++ + ++/+++ Usability +++ +++ ++ +++/+

Page 375: genetic_POPULATIONS.doc

generating a relatively small number of points, it is practical to use the systematic

Page 376: genetic_POPULATIONS.doc

sampling process only in small dimensions. Otherwise, the proportion of genuine

Page 377: genetic_POPULATIONS.doc

nonaligned systematic sample points is small. This is a strong limitation in the area of

Page 378: genetic_POPULATIONS.doc

optimization.

Page 379: genetic_POPULATIONS.doc

4.5 Summary of features

Page 380: genetic_POPULATIONS.doc

An ideal generator for our purposes should generate well-distributed, genetically diverse points in n-dimensions, and it should be relatively fast and easy to use. In Table 2, we summarize the evaluation made in this section. The evaluation of differ- ent point generators has been marked using plus signs (+): the more plus signs the better the score. In case of a notable difference in how well a generator works for

Page 381: genetic_POPULATIONS.doc

problems with small and large dimensions (here small denotes less than, say, five), the occasions are scored separately (small/large). For example, the nonaligned systematic sampling scores +++/+ in coverage,whichmeans that itworkswell insmall dimensions

Page 382: genetic_POPULATIONS.doc

and poorly in large dimensions.

Page 383: genetic_POPULATIONS.doc

Note that only coverage and genetic diversitymay have a direct influence on objec-

Page 384: genetic_POPULATIONS.doc

tive function values inoptimization since they are the properties of the points whereas speed and usability are properties of the generators. Considering just the properties

Page 385: genetic_POPULATIONS.doc

of the points we notice in Table 2 that pseudo random points have good genetic diver- sity, but the worst coverage, and the SSI process produces points with good coverage, but only average genetic diversity. The properties of the Niederreiter quasi random points settle somewhere between pseudo and the SSI process. In the further analysis, we pay less attention to the nonaligned systematic sampling since it is not applica- ble for problems with several variables. To find out the effects of the coverage and genetic diversity, we concentrate on the pseudo random number generator and the

Page 386: genetic_POPULATIONS.doc

SSI process.

Page 387: genetic_POPULATIONS.doc

5 Experimentations

Page 388: genetic_POPULATIONS.doc

We test the influence of initial populations computationally by using different point

Page 389: genetic_POPULATIONS.doc

generators and a largenumber of difficult test problems from the literature. The influ-

Page 390: genetic_POPULATIONS.doc

ence of the different initial populations is analyzed after 10 and 20 generations and after the execution of the whole algorithm. The genetic algorithm used is presented

Page 391: genetic_POPULATIONS.doc

in Section 2.

Page 392: genetic_POPULATIONS.doc

5.1 Test settings

Page 393: genetic_POPULATIONS.doc

The test runs wereperformed on anHP9000/J5600computer.We solved a test suite of

Page 394: genetic_POPULATIONS.doc

52problems using geneticalgorithms described in Section 2 with theparameter values

Page 395: genetic_POPULATIONS.doc

Tab le 2 Summary of generator properties

Properties (small/large) Pseudo Nieder SSI Nonalig

Page 396: genetic_POPULATIONS.doc

Genetic diversity

Page 397: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 422

Value

201 21 3 0.8 0.1

Steps 100 1 0 −7

# Name Ref.

1–3 2, 5, 10 [0, π] [13] 4–7 Rastrigin [13] 8–11 Schwefel [13] 12 2 [13] 13–17 [13] 18–22 [13] 23 Easom 2 [13] 24 Levy 4 [−10, 10] [40] 25–27 Levy 5, 6, 7 [−5, 5] [40] 28 P8 3 [−10, 10] [7] 29 P16 5 [−5, 5] [7] 30 Hansen 2 [−10, 10] [40] 31–34 Corona 4, 6, 10, 20 [10] 35–38 Katsuura 4, 6, 10, 20 [−1, 1] [10] 39–42 5, 5, 10, 10 [0, 10] [41, 13] 43–46 3, 4, 10, 20 [−20, 20] [42] 47–49 Shekel 4, 4, 4 [0, 10] [40] 50–52 2, 5, 10 [0, π] [41]

Page 398: genetic_POPULATIONS.doc

Tab le 3 Parameter values for genetic algorithms

Parameter

Page 399: genetic_POPULATIONS.doc

Population size

Page 400: genetic_POPULATIONS.doc

Elitism size

Page 401: genetic_POPULATIONS.doc

Tournament size

Page 402: genetic_POPULATIONS.doc

Crossover rate

Page 403: genetic_POPULATIONS.doc

Mutation rate

Page 404: genetic_POPULATIONS.doc

Max generations 10, 20 and 2000

Page 405: genetic_POPULATIONS.doc

To le ran ce

Page 406: genetic_POPULATIONS.doc

given in Table 3. The only difference between the variants of a genetic algorithm used was the way the initial population was generated. Each problem was solved 10 times with each algorithm, when the algorithms were let run until the stopping criteria were

Page 407: genetic_POPULATIONS.doc

satisfied, and 100times when thealgorithms were stopped prematurely after 10 and 20 iterations. In each run, we recorded the best objective function value in the last pop- ulation and study them in what follows. The names of the function families, number of variables, box constraints used and references are collected in Table 4. If only one interval is given for the box constraints in Table 4, then each variable was restricted to the same interval. The test problems are the same as used in [24]. They are divided into two problem sets according to the number of variables so that problems with 10

Page 408: genetic_POPULATIONS.doc

or less variables are here called small and others large.

Page 409: genetic_POPULATIONS.doc

Originally, initial populationswere generated in altogether 21 differentways includ-

Page 410: genetic_POPULATIONS.doc

ing 13 variants of the SSI process, 5 quasi random number generators, a nonaligned systematic sampling and a pseudo random number generator. The variants of the SSI process differed in the random number generator used when generating the trial points and in the metric used when computing the probabilistic maximal minimum distance PMMD. The quasi random number generators tested were the Niederreiter,

Page 411: genetic_POPULATIONS.doc

Tab le 4 Summary of the test problems

Page 412: genetic_POPULATIONS.doc

Dimensions Box constraints

Page 413: genetic_POPULATIONS.doc

Michalewicz

Page 414: genetic_POPULATIONS.doc

6, 10, 20, 30 [−600, 400] 6, 10, 20, 50 [−500, 500]

Page 415: genetic_POPULATIONS.doc

Branin rcos [−5, 10]×[0, 15] Griewangk 2, 6, 10, 20, 50 [−700, 500] Ackley’s Path 2, 6, 10, 20, 30 [−30.768, 38.768]

Page 416: genetic_POPULATIONS.doc

[−100, 100]

Page 417: genetic_POPULATIONS.doc

[−900, 1100]

Page 418: genetic_POPULATIONS.doc

Langerman

Page 419: genetic_POPULATIONS.doc

Funtion 10

Page 420: genetic_POPULATIONS.doc

Epistatic Michalewicz

Page 421: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 423

4).

f (x)

problems

Pseudo Nieder SSI Nonalig

f (x)Small −186.7 −186.7 −186.8 −186.6 f (x)Large −1896.0 −2205.4 −1680.2 −2295.9

0.34 0.28 0.37 0.58 885.9 355.3 162.5 1697.9

Page 422: genetic_POPULATIONS.doc

Faure and Halton generator and two variants of the Sobol’ generator. The numerical

Page 423: genetic_POPULATIONS.doc

results of the genetic algorithm with quasi random initial populations are reported in [23], where the Niederreiter generator performed the best. Here, we report the results of the genetic algorithm using initial populations generated by the good rep-

Page 424: genetic_POPULATIONS.doc

resentatives from different types of point generators (see the beginning of Section

Page 425: genetic_POPULATIONS.doc

5.2 Numerical results

Page 426: genetic_POPULATIONS.doc

In this subsection, we present the results of the experiments described above. In what

Page 427: genetic_POPULATIONS.doc

follows, we consider only the best objective function values that the algorithm has

Page 428: genetic_POPULATIONS.doc

found (in the last population) and pay attention to the average and variance of these

Page 429: genetic_POPULATIONS.doc

values in the repeated runs.

Page 430: genetic_POPULATIONS.doc

Let us first consider those tests where the geneticalgorithm was run until a stopping

Page 431: genetic_POPULATIONS.doc

criterion was satisfied. Table 5 shows the average final objective function values

Page 432: genetic_POPULATIONS.doc

and the average standard deviations σ(f )for small and largeproblems.As we can see,

Page 433: genetic_POPULATIONS.doc

for smallproblems, in the average objective function values, there are no differences in three significant digits. However, for large problems the differences in average objec- tive function values are quite large. Based solely on the average objective function

Page 434: genetic_POPULATIONS.doc

values in Table 5 one could draw the conclusion that nonaligned systematic samples

Page 435: genetic_POPULATIONS.doc

and Niederreiter quasi random points are clearly the most suitable for initial popula-

Page 436: genetic_POPULATIONS.doc

tions. This, however, is not completely true as our further analysis reveals, when we

Page 437: genetic_POPULATIONS.doc

consider also the variances.

Page 438: genetic_POPULATIONS.doc

Before any statistical tests, we make an observation on the results in Table 5. As

Page 439: genetic_POPULATIONS.doc

mentioned in Section 4.2, nonaligned systematic sampling reduces to pseudo random sampling for large problems. However, the average objective function value and the

Page 440: genetic_POPULATIONS.doc

average standard deviation of those algorithms differ quiteremarkably for large prob-

Page 441: genetic_POPULATIONS.doc

lems. This indicates that the average values are not good measures in the comparison.

Page 442: genetic_POPULATIONS.doc

We applied analysis of variance (ANOVA) to the final objective function values to

Page 443: genetic_POPULATIONS.doc

find out whether the differences were statistically significant. The analysis of variance

Page 444: genetic_POPULATIONS.doc

indicated that only in 4 out of 52 test problems there were statistically significant

Page 445: genetic_POPULATIONS.doc

differences (with95% confidence).We call these four problems critical problems,and

Page 446: genetic_POPULATIONS.doc

they are the problems 40, 49, 51 and 52 in Table 4. Surprisingly, all the critical prob-

Page 447: genetic_POPULATIONS.doc

lems were small (problems with 10 or less variables). This can be explained by the small variance for these problems. The fact that there was no statistically significant

Page 448: genetic_POPULATIONS.doc

difference in the objective function values for any of the large problems is explained by their very large variances. The variances were the largest for the 20-dimensional

Page 449: genetic_POPULATIONS.doc

Katsuura problem for which the variances were 8,473, 2,060, 509 and 17,100 for the pseudo, Niederreiter, SSI and nonaligned, respectively. The 20-dimensional Katsuura

Page 450: genetic_POPULATIONS.doc

Tab le 5 Average objective

Page 451: genetic_POPULATIONS.doc

function values and standard

Page 452: genetic_POPULATIONS.doc

deviations for small and large

Page 453: genetic_POPULATIONS.doc

σ(f )Small

Page 454: genetic_POPULATIONS.doc

σ(f )Large

Page 455: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 424

in one.

sf =f − f min

f max − f min

,

-0.8

-0.6

Ps Ni SSI No

-14

-10

-6

-2

Ps Ni SSI No -4.7

-4.6

-4.5

Ps Ni SSI No

-8.8

Ps Ni SSI No -1

Problem 51 (P=0.001)

-9.6

-9.2

Page 456: genetic_POPULATIONS.doc

problem biased the average values in Table 5 and was the main cause of differences

Page 457: genetic_POPULATIONS.doc

for large problems.

Page 458: genetic_POPULATIONS.doc

Next, we study the critical problems more closely. Figure 9 shows the average

Page 459: genetic_POPULATIONS.doc

objective function values for the four critical problems. Ps, Ni, SSI and No stand for

Page 460: genetic_POPULATIONS.doc

the pseudo, Niederreiter, SSI process and nonaligned systematic sampling, respec-

Page 461: genetic_POPULATIONS.doc

tively, and the whiskers illustrate the standard deviations. In addition, the P values related to the F test of ANOVA [28] are given in the parenthesis subsequent to the problem number. The four plots reveal that there is no overall dominance between

Page 462: genetic_POPULATIONS.doc

the geneticalgorithm variants, however the variant applying quasirandom points per-

Page 463: genetic_POPULATIONS.doc

formed best in 3 out of 4 cases and also points generated by the SSI process performed better than pseudo random points in 3 out of 4 instances. Nonaligned systematic sam-

Page 464: genetic_POPULATIONS.doc

pling performed very similarly to pseudo random points in three cases and was better

Page 465: genetic_POPULATIONS.doc

So far we have studied the differences in objective function values after running

Page 466: genetic_POPULATIONS.doc

the whole algorithm and there has been only minor differences. Next, we consider also the situations, where the algorithm is stopped prematurely after 10 and 20 gener-

Page 467: genetic_POPULATIONS.doc

ations. Earlier, we noticed difficulties in interpreting values that were not scaled: one

Page 468: genetic_POPULATIONS.doc

large value could strongly bias the results. Therefore, we now define a more reliable measure of performance by scaling the objective function values to the range from 0

Page 469: genetic_POPULATIONS.doc

to 1. Scaled average objective function value sf is defined as follows

Page 470: genetic_POPULATIONS.doc

Problem 40 (P=0.007)

Problem 49 (P=0.040)

Page 471: genetic_POPULATIONS.doc

Problem 52 (P=0.043)

Page 472: genetic_POPULATIONS.doc

Fig. 9 Average objective function values and standard deviations

Page 473: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 425

min

max

max

min

min

max

f max − f min

results.

0.90 0.91 0.85 0.88 0.9 0.87

all 0.52 0.38 0.84 0.85 0.91 0.90

all 0.45 0.61

Page 474: genetic_POPULATIONS.doc

where f is the best objective function value known (or found after running all the

Page 475: genetic_POPULATIONS.doc

four whole algorithms), and f is the worst of the best objective function value

Page 476: genetic_POPULATIONS.doc

found during the repeated runs (either for a fixed number of generations or till a

Page 477: genetic_POPULATIONS.doc

stopping criterion). It is worth noting that f changes each time the stopping criteria are changed, but f typically remains the same (at least if it is known). Finally, a

Page 478: genetic_POPULATIONS.doc

bar above a function symbol denotes the average value over all the considered test problems (small or large), where the difference between (unscaled values of) f and f was more than 0.1. Table 6 shows the scaled average objective function values for

Page 479: genetic_POPULATIONS.doc

small and large problems.

Page 480: genetic_POPULATIONS.doc

We remind again that for large dimensional problems the nonaligned systematic

Page 481: genetic_POPULATIONS.doc

sampling points and pseudo random points are essentially the same, and their results

Page 482: genetic_POPULATIONS.doc

are here reported only for the sake of completeness. We now notice in Table 6 that

Page 483: genetic_POPULATIONS.doc

pseudo and nonaligned systematic sampling receive similar values, which gives indi-

Page 484: genetic_POPULATIONS.doc

cation that the values are more reliable after scaling.

Page 485: genetic_POPULATIONS.doc

The scaled average objective function value sf becomes unstable near theoptimum

Page 486: genetic_POPULATIONS.doc

since the divisor approaches zero and, therefore, the values for 10 and

Page 487: genetic_POPULATIONS.doc

20 generations are more reliable than the values after running the whole algorithm. Nevertheless, Table 6 shows an interesting trend: the genetic algorithm variants that converge relatively fast in the first generations are not necessarily the ones obtain-

Page 488: genetic_POPULATIONS.doc

ing the best final results. This is particularly true for the SSI process, which had the

Page 489: genetic_POPULATIONS.doc

best empty space statistic values in Section 4 and in the numerical results in Table 6 receives worst values during the first generations, but good values at the end. Pseudo random points, on the other hand, had the best genetic diversity, and, in Table 6, they receive good values during the first iterations. The Niederreiter quasi random points had above average genetic diversity and above average coverage and their performance in numerical tests in Table 6 is above average except for one instance.

Page 490: genetic_POPULATIONS.doc

Despite the observations made above, the differences in the results shown in Table 6

Page 491: genetic_POPULATIONS.doc

are debatable.

Page 492: genetic_POPULATIONS.doc

We applied ANOVA also for the intermediate results after 10 and 20 generations.

Page 493: genetic_POPULATIONS.doc

Again, we call those problems critical, where there were statistical differences in the objective function values. However, we now exclude the problems, where the differ- ences in objective function values were less than 0.1. There were all together 38 critical problems after 10 generations and 9 after 20 generations. In Table 7, we report the number of critical problems for which each algorithm had the best average interme- diate objective function value. The differences are not large, but pseudo random (and

Page 494: genetic_POPULATIONS.doc

nonaligned systematic sampling) initial population seem to provide best intermediate

Page 495: genetic_POPULATIONS.doc

Tab le 6 Scaled average objective function values after

Generations Pseudo Nieder SSI Nonalig 10 and 20 generations and after running the whole algorithm sf (x)Small 10

sf (x)Small 20 0.95 0.89

Page 496: genetic_POPULATIONS.doc

sf (x)Small 0.43 0.57 sf (x)Large 10 0.99 0.84 sf (x)Large 20 0.99 0.87 sf (x)Large 0.38 0.42

Page 497: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 426

Total

10 9 9 8 12 38 20 5 1 0 3 9

Total

10 18 11 8 9 46 20 12 11 6 15 44 Final 5 9 3 2 19

Page 498: genetic_POPULATIONS.doc

Tab le 7 The number of problems with best average

Generations Pseudo Nieder SSI Nonalig

Page 499: genetic_POPULATIONS.doc

objective function values for

Page 500: genetic_POPULATIONS.doc

critical problems

Page 501: genetic_POPULATIONS.doc

Before summarizing the results, let us yet consider the magnitudes of the variances

Page 502: genetic_POPULATIONS.doc

in the best objective function values separately from the actual objective function

Page 503: genetic_POPULATIONS.doc

values. In Table 8 are reported the number of test problems for which each algorithm had the smallest variance after 10 and 20 generations and after running the whole algorithms. We have excluded the problems where the differences in variances were less than 0.1. The total number of problems considered is reported in the last column of Table 8. It is noteworthy that the variant of genetic algorithm using pseudo random numbers has most often the smallest variance for the intermediate solutions, which

Page 504: genetic_POPULATIONS.doc

was not expected.

Page 505: genetic_POPULATIONS.doc

5.3 Summary of numerical results

Page 506: genetic_POPULATIONS.doc

Here, we summarize the analysis made above.When considering final and intermedi-

Page 507: genetic_POPULATIONS.doc

ate average objective function values, the analysis of variance indicated four critical

Page 508: genetic_POPULATIONS.doc

problems for the final results and many more for the intermediate results where there

Page 509: genetic_POPULATIONS.doc

were significant differences in the best objective function values found. None of the methods clearlydominated the others, but in the scaled objective function values there

Page 510: genetic_POPULATIONS.doc

was a noticeable trend, namely that pseudo random initial populations provided on the average good intermediate results, but not so good final results. The converse was true for the SSI process and quasi random points with better uniform coverage. Ear-

Page 511: genetic_POPULATIONS.doc

lier we noticed that some clusterizations were harmful and some beneficial. Hence,

Page 512: genetic_POPULATIONS.doc

genetic algorithms using pseudo random numbers and the variants where the initial

Page 513: genetic_POPULATIONS.doc

populations have better uniform coverage performed in a slightly different way. Nev- ertheless, the differences, in general, were debatable and no strong conclusions could

Page 514: genetic_POPULATIONS.doc

be made.

Page 515: genetic_POPULATIONS.doc

As far as the magnitudes of variances in the intermediate results are concerned,

Page 516: genetic_POPULATIONS.doc

in most of the cases, the variance was the smallest for the genetic algorithm using

Page 517: genetic_POPULATIONS.doc

pseudo random numbers. There may be both harmful and beneficial clusterizations

Page 518: genetic_POPULATIONS.doc

as seen in Figs. 1 and 2. Harmful clusterizations lead to inferior and beneficial lead to superior objective function values compared to initial populations without clustering. One could have expected this to imply more significant differences in the magnitudes

Page 519: genetic_POPULATIONS.doc

of variances in the early generations but this did not happen. However, again, none

Page 520: genetic_POPULATIONS.doc

of the algorithms clearly dominated the others and the differences were not large.

Page 521: genetic_POPULATIONS.doc

Finally, there were no significant differences in the magnitudes of variances related

Page 522: genetic_POPULATIONS.doc

to final objective function values when applying clustered or unclustered initial popu-

Page 523: genetic_POPULATIONS.doc

Tab le 8 Number of test problems when the variance

Generations Pseudo Nieder SSI Nonalig

Page 524: genetic_POPULATIONS.doc

was smallest

Page 525: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 427

b) a)

Page 526: genetic_POPULATIONS.doc

lations.This could be assumed keeping inmind the observation that geneticalgorithms

Page 527: genetic_POPULATIONS.doc

areMarkov chains and, hence, converge independently of the initial population.How-

Page 528: genetic_POPULATIONS.doc

ever, the conclusion is not trivial, since actually theMarkov chain property only guar-

Page 529: genetic_POPULATIONS.doc

antees convergence when thenumber of generations approaches infinity. The last row

Page 530: genetic_POPULATIONS.doc

in Table 8 shows that the geneticalgorithm applying pseudorandom initial population

Page 531: genetic_POPULATIONS.doc

has more often smallest variance than the one applying the SSI process, but less often than the one applying quasi random sequences. Hence, the clusterization of the initial population did not seem to have strong effect on the variance of the final objective

Page 532: genetic_POPULATIONS.doc

function values.

Page 533: genetic_POPULATIONS.doc

Despite some differences, the results indicate that pseudo random points perform

Page 534: genetic_POPULATIONS.doc

rather well when used in initial populations of genetic algorithms. Moreover, if the

Page 535: genetic_POPULATIONS.doc

minor differences found in this research can be generalized, then they speak for

Page 536: genetic_POPULATIONS.doc

pseudo random points in cases when the algorithm is stopped after relatively small

Page 537: genetic_POPULATIONS.doc

number of generations, which may be the case with some computationally expensive

Page 538: genetic_POPULATIONS.doc

real life problems.

Page 539: genetic_POPULATIONS.doc

6 Discussion

Page 540: genetic_POPULATIONS.doc

In this section, we consider fourproblematic issues that we have touched earlier. First, we point out that dimensionality plays a large role in this research. The initial popula- tions become sparse when the optimization problem has more than few variables. A

Page 541: genetic_POPULATIONS.doc

simple example of this can be seen with unit hyper cubes in three and two dimensions.

Page 542: genetic_POPULATIONS.doc

In three dimensions, eight point can be located in the corners of the cube so that

Page 543: genetic_POPULATIONS.doc

their maximal minimum distance is always one. But locating eight points in the line

Page 544: genetic_POPULATIONS.doc

segment of length one means that the maximal minimum distance between the points is 1/7. The sparsity in higher dimensions means that solutions in the initial population are likely to be relatively far from the global optimum no matter what uniform distri- bution pattern is used. Moreover, in sparse populations, when the minimum distance

Page 545: genetic_POPULATIONS.doc

between two adjacent points in the population is maximized, that is, when the points

Page 546: genetic_POPULATIONS.doc

in the population try tomaximally avoid each other, then the population is likely to be

Page 547: genetic_POPULATIONS.doc

concentrated to the borders of the feasible region, whichmay not always be desirable.

Page 548: genetic_POPULATIONS.doc

This phenomenon is typical of the SSI process and could also be seen in the example

Page 549: genetic_POPULATIONS.doc

mentioned above.

Page 550: genetic_POPULATIONS.doc

The second issue is of more technical nature and concerns the SSI process. The SSI

Page 551: genetic_POPULATIONS.doc

process can produce points that are either genetically diverse or have good uniform coverage, depending on the distance parameter PMMD. We optimized the coverage, but did not test values forPMMD that would compromise between the two conflicting objectives. Smaller values for would also make the process considerably faster. For example, in Fig. 10a we have seven point in an interval of length one with = 1/7.

Page 552: genetic_POPULATIONS.doc

We know that one more point can be included but how many random points should

Page 553: genetic_POPULATIONS.doc

be generated before the right location is found?

Page 554: genetic_POPULATIONS.doc

When optimizing the coverage for the SSI process the automatic determination of

Page 555: genetic_POPULATIONS.doc

a maximal minimum distance PMMD is problematic. We did not find good theoret-

Page 556: genetic_POPULATIONS.doc

ical estimates for PMMD and the theoretical estimates we tested did not work well

Page 557: genetic_POPULATIONS.doc

Fig. 10 Difficulties with speed

Page 558: genetic_POPULATIONS.doc

and approximation of

Page 559: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 428

Page 560: genetic_POPULATIONS.doc

in practice. Hence, we used experimental estimates. In our current implementation,

Page 561: genetic_POPULATIONS.doc

we have numerically solved maximal values of PMMD for some number of points

Page 562: genetic_POPULATIONS.doc

in some dimensions, and we approximate maximal PMMD elsewhere using interpo-

Page 563: genetic_POPULATIONS.doc

lation. Examples of values used for in our experiments for different numbers of variables (with population size 200) include 0.032 (for n = 2), 0.23 (for n = 5), 0.49 (for n = 10), 0.64 (for n = 20), 0.79 (for n = 50) and 0.84 (for n = 100). Some further research is needed before the SSI process with maximal values for PMMD can be

Page 564: genetic_POPULATIONS.doc

used in practical applications.Anexample of the difficulty of approximating the value

Page 565: genetic_POPULATIONS.doc

of is given in Fig. 10b. Let us assume that we have fixed = 1 / 7 a nd t h e d istance

Page 566: genetic_POPULATIONS.doc

between the end points of the interval and outermost points is 1 / 8 a nd t h e d istance

Page 567: genetic_POPULATIONS.doc

between intermediate points is 2/8. Here we have only four points (denoted by black

Page 568: genetic_POPULATIONS.doc

dots in the figure) and cannot fit any more points even though eight points could be

Page 569: genetic_POPULATIONS.doc

fit in the interval in the optimal case (denoted by circles in the figure). In other words, it is very hard to determine the maximal minimum distance a priori when points are

Page 570: genetic_POPULATIONS.doc

generated randomly.

Page 571: genetic_POPULATIONS.doc

The third issue concerns the speed of point generators.We point out that the com-

Page 572: genetic_POPULATIONS.doc

putational complexity of an algorithm may depend strongly on the implementation,

Page 573: genetic_POPULATIONS.doc

see, e.g., [4], where bit-operations are considered. Thus, the results on the speed of

Page 574: genetic_POPULATIONS.doc

the generators can be considered only suggestive.

Page 575: genetic_POPULATIONS.doc

The last issue concerns the testing of point generators. There exists a large variety

Page 576: genetic_POPULATIONS.doc

of dynamic statistical tests for the independence of points generated by point gener- ators [14]. These tests are called goodness-of-fit tests. One well-established battery of goodness-to-fit tests is called DIEHARD and the source code is available at [8]. However, the goodness-to-fit tests are designed to test the independence of large

Page 577: genetic_POPULATIONS.doc

numbers of points. For example, DIEHARD requires a binary file of a size 10–11 megabytes to evaluate a generator. Since we are only interested in a few hundred

Page 578: genetic_POPULATIONS.doc

numbers at the beginning of the sequence, we used the Ripley’s K-function instead of

Page 579: genetic_POPULATIONS.doc

the goodness-to-fit tests.

Page 580: genetic_POPULATIONS.doc

7 Conclusions and future research

Page 581: genetic_POPULATIONS.doc

In this paper, we have tested different initial populations for a real coded genetic

Page 582: genetic_POPULATIONS.doc

algorithm. Traditionally, pseudo random numbers are used for generating initial pop-

Page 583: genetic_POPULATIONS.doc

ulation, and our motivation has been to study and initiate discussion on whether their

Page 584: genetic_POPULATIONS.doc

use is justified.

Page 585: genetic_POPULATIONS.doc

We have shown with asimple, academic,example that initial populationsmay have

Page 586: genetic_POPULATIONS.doc

a significant effect on the best objective function value over several generations. Then

Page 587: genetic_POPULATIONS.doc

we have concentrated on studying different realistic ways to generate initial popu-

Page 588: genetic_POPULATIONS.doc

lations for a case with no a priori information on the location of the global minima. We have briefly summarized the basic properties of a pseudo and a quasi random sequence generator, the SSI process and the nonaligned systematic sampling and applied them. We have discussed their properties including genetic diversity and a gooduniform coverage for the points as well as speed and usability for the generators. In thenumerical tests with geneticalgorithms, we have used a test suite of 52 functions

Page 589: genetic_POPULATIONS.doc

from the literature.We have studied the effects of the different initial populations on

Page 590: genetic_POPULATIONS.doc

the best objective function values after 10 and 20 generations and after running the

Page 591: genetic_POPULATIONS.doc

whole algorithm.

Page 592: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 429

102020.

Page 593: genetic_POPULATIONS.doc

There were differences in the coverage and genetic diversity of the tested point

Page 594: genetic_POPULATIONS.doc

sets. The SSI process has a good uniform coverage, but only average genetic diversity, and for pseudo random points it is vice versa. The Niederreiter quasi random points

Page 595: genetic_POPULATIONS.doc

have both above average coverage and above average genetic diversity.

Page 596: genetic_POPULATIONS.doc

In the numerical experiments with the genetic algorithms, there was a trend—

Page 597: genetic_POPULATIONS.doc

although weak—showing that the versions of genetic algorithm with good genetic diversity converged fast during the first generations, but did not obtain the best final objective function values on the average. The converse was true for the versions with good uniform coverage. However, the differences were not so large that any strong

Page 598: genetic_POPULATIONS.doc

conclusion could be drawn.

Page 599: genetic_POPULATIONS.doc

With respect to the speed and usability of the point generators the results show that

Page 600: genetic_POPULATIONS.doc

pseudo and quasi random sequence generators are fast and easy to use. Both the SSI process and the nonaligned systematic sampling require more developing and testing.

Page 601: genetic_POPULATIONS.doc

We conclude that, based on our research, the traditional way to generate initial

Page 602: genetic_POPULATIONS.doc

populations of genetic algorithms using pseudo random number generator was not worse than the others. It is particularly well suited to cases where the algorithm must be stopped prematurely; which may happen with computationally expensive real

Page 603: genetic_POPULATIONS.doc

life problems. However, there are also good alternative ways such as quasi random

Page 604: genetic_POPULATIONS.doc

sequences and the SSI process, which have advantages in certain cases, especially if the goodness of the final solution is valued higher that the speed of convergence

Page 605: genetic_POPULATIONS.doc

during the first generations.

Page 606: genetic_POPULATIONS.doc

Our findings show that paying attention to the initial population may have an

Page 607: genetic_POPULATIONS.doc

effect on the success of the genetic algorithm and further research and discussion is encouraged. The topics for future research include, firstly, to study different ini- tial populations for specific types of problems and with different genetic algorithm

Page 608: genetic_POPULATIONS.doc

parameters like population size and maximum number of generations. Secondly, to

Page 609: genetic_POPULATIONS.doc

study more closely the problem of dimensionality discussed in Section 6, and thirdly, to further develop the SSI implementation and to discover good theoretical estimates

Page 610: genetic_POPULATIONS.doc

for the maximal minimum distance PMMD.

Page 611: genetic_POPULATIONS.doc

Acknowledgements The authors thank doctors Salme Kärkkäinen andMarkoM.Mäkelä for useful discussions and professor Michael Mascagni for providing the Niederreiter generator. This research was supported by the Jenny and AnttiWihuri Foundation and the Academy of Finland, grant number

Page 612: genetic_POPULATIONS.doc

Appendix A

Page 613: genetic_POPULATIONS.doc

This appendix is designed to give a short overview to different pseudorandomnumber and quasi random sequence generators. In the following definitions we use (y)modM

Page 614: genetic_POPULATIONS.doc

to denote y modulo M.

Page 615: genetic_POPULATIONS.doc

A.1 Some pseudo random number generators

Page 616: genetic_POPULATIONS.doc

Traditionally, pseudo random number generators produce sequences of scalars. If vectors are needed, then they are usually formed by taking the first n scalars for the first n-dimensional vector, the next n scalars for the second vector and so forth. According to [30], it would be preferable to generate pseudo random vectors directly. Here, we present traditional pseudo random number generators and one vector

Page 617: genetic_POPULATIONS.doc

generator.

Page 618: genetic_POPULATIONS.doc

i

0 = y0 i

)

congruential generators with a = 0 a nd a =0. Then a generator with a =0 is called a

i = (a i−1

2 2 2

1

J Glob Optim (2007) 37:405–436

2 ˆi ˆy = (a ˆi

430

j

0

i

i ˆ i

Linear y 1 y 2 )modM

y i = (a 1 (y i−1 2 y +a 3 )modM

y i = (a 1 y 1 i−1 2

y i−1

y i +y i−1

x = k i y i

M

1

)modM 2

3 ̂ y −1 )modM 3

x y i i = ( M 1

ˆyi + M 2

ˆi ˆy + M 3 )mod 1.

2

j i

Page 619: genetic_POPULATIONS.doc

We classify pseudo random number generators to two main categories: congruen-

Page 620: genetic_POPULATIONS.doc

tial and recursive generators.We call amethod congruential, if it usesmodulo and only the previous iteration value, andwe call it recursive, if it uses values from several previ-

Page 621: genetic_POPULATIONS.doc

ous iterations. In addition, we present a feedback shift register generator and a vector generator called SQRT-generator, which do not fall into the two main categories. For another classification, see, e.g., [39]. The classifications are always somewhat vague.

Page 622: genetic_POPULATIONS.doc

Feedback shift register generator also uses values from several previous iterations,

Page 623: genetic_POPULATIONS.doc

but its construction differs significantly from recursive generators, and therefore, we

Page 624: genetic_POPULATIONS.doc

put it in its own class. For more detailed information about pseudo random number

Page 625: genetic_POPULATIONS.doc

generation we refer to [14] and to references therein.

Page 626: genetic_POPULATIONS.doc

A.1.1 Congruential generators

Page 627: genetic_POPULATIONS.doc

Next, we present five congruential generators. They are called linear, quadratic, inver-

Page 628: genetic_POPULATIONS.doc

sive, additive and parallel linear congruential generators. The name congruential

Page 629: genetic_POPULATIONS.doc

comes from the use of modulo. For example, in modulo 5, the numbers 4 and 9 are called congruent. In what follows,M is themodulo and a ’s are prescribed integers.On

Page 630: genetic_POPULATIONS.doc

the ith iteration, the generators first produce an integer y i ∈ [ 0,M)and then arandom number x i ∈ [ 0, 1)is obtained by a division xi = yi /M unless another formula is given.

Page 631: genetic_POPULATIONS.doc

The seed y is a large prescribed integer. An additive congruential generator (of kth

Page 632: genetic_POPULATIONS.doc

order) [45], to be described below, requires k 0 +1 seeds 0 ≤ yj < M, j = 0, ... , k ,a nd

Page 633: genetic_POPULATIONS.doc

the parallel linear generator includes three separate linear generators denoted by y ,

Page 634: genetic_POPULATIONS.doc

ˆy and ˆy .

Page 635: genetic_POPULATIONS.doc

i = (a i−1 +a

Page 636: genetic_POPULATIONS.doc

Quadratic

2 +a i−1

Inversive

+a )modM

Page 637: genetic_POPULATIONS.doc

Additive

Page 638: genetic_POPULATIONS.doc

m = (ym−1 m )modM,m = 1, ..., k

Page 639: genetic_POPULATIONS.doc

Parallel linear y

y )modM

Page 640: genetic_POPULATIONS.doc

ˆyi = (a ˆyi−1

Page 641: genetic_POPULATIONS.doc

In all of the above generators i = 1, 2, ... Often, a distinction is made between linear

multiplicative linear congruential generator and a generator with a =0 a mixed linear

Page 642: genetic_POPULATIONS.doc

congruential generator.

Page 643: genetic_POPULATIONS.doc

A.1.2 Recursive generators

Page 644: genetic_POPULATIONS.doc

Generators using two or more previous random numbers when generating a new number in the sequence are here called recursive generators. In what follows, a , c , s,

Page 645: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

−y i− r −c i

of zeros and ones.

p ∈ [ 0,M) can be selected arbitrarily.

431

i

i = yi

y 1 y k y i−k )modM

y )modM

y +y i− r +c i

c =i

i = (yi−s

1, y

c =i 1, if y i−s < 0

y 1 y +a i−r 2 y +c i i is i−1 i

Again i

j

i j

p 1

p−1 +· · · +a −1z +a ) , ( 1) p p

i i

n p

n−p p

n−p+1 +· · · +a α ) m o d M . ( 2) 1 n−1

1

Page 646: genetic_POPULATIONS.doc

and r are prescribed integers s < r, j

= 1, ... , k , a nd ci is called the carry operator on

Page 647: genetic_POPULATIONS.doc

the ith iteration. The determination of c is omitted for the multiply-with-carry-gener- ator and also for the add-with-carry and substract-with-borrow generators usingmore

Page 648: genetic_POPULATIONS.doc

that two previous values. Again, the pseudo random number x i ∈ [ 0, 1)is obtained by a division x /M. For more details about the carry operator, see, e.g., [6]. Next, we

Page 649: genetic_POPULATIONS.doc

present five recursive generators.

Page 650: genetic_POPULATIONS.doc

Multiplicative recursive

i = (a i−1 +· · · +a

Lagged Fibonacci

i = (yi−r −y i− s

Add −with −carry ) m o d M , w here

Page 651: genetic_POPULATIONS.doc

i−s +y i− r +c i−1 ≤ M

Page 652: genetic_POPULATIONS.doc

0, otherwise

Page 653: genetic_POPULATIONS.doc

Substract

−with −borrow yi = (yi−s −y i− r −c i ) m o d M , w here

Page 654: genetic_POPULATIONS.doc

0, otherwise

Page 655: genetic_POPULATIONS.doc

Multiply −with −carry i = (a i−s )modM ,w here c

Page 656: genetic_POPULATIONS.doc

computed using c and previous values of y .

Page 657: genetic_POPULATIONS.doc

= 1, 2, ... and y0

is the seed. The number of previous values of y used in the

Page 658: genetic_POPULATIONS.doc

methods described above may alter. In addition, the lagged Fibonacci generator may

Page 659: genetic_POPULATIONS.doc

generate new values also as a sum or a product of the previous values.

Page 660: genetic_POPULATIONS.doc

A.1.3 Feedback shift register generator

Page 661: genetic_POPULATIONS.doc

Feedback shift register generator uses modulo in a different manner compared to the congruential and recursive generators. Modulo M is not a large integer, but very

Page 662: genetic_POPULATIONS.doc

often M = 2 in which case the generator produces a sequence α

Page 663: genetic_POPULATIONS.doc

The sequence is then cut into subsequences with an appropriate length. These subse- quences correspond to y ’s. To generate the sequence α , first a primitivepolynomial

Page 664: genetic_POPULATIONS.doc

q (z ) of order p is selected

Page 665: genetic_POPULATIONS.doc

q(z) = z −(a z

Page 666: genetic_POPULATIONS.doc

where a ∈{0, 1, ... ,M −1}, i = 1, ... , p. Then using the coefficients a , t he nth num-

Page 667: genetic_POPULATIONS.doc

ber in the sequence is generated using the formula

Page 668: genetic_POPULATIONS.doc

α = (a α

+a − 1α

Page 669: genetic_POPULATIONS.doc

The initial numbers α , ... , α

Page 670: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

i √p , i = 1, ... , n. Then the n-dimensional SQRT sequence

432

i i

i 1 n

E =n

i=1

a i

b d i ,

a +1 i

b d i ,

i i i d i i

n ⊂ Rn d i

1 2 1 2

t

t−m 2

m

τ

1

1

Page 671: genetic_POPULATIONS.doc

A.1.4 SQRT sequence

Page 672: genetic_POPULATIONS.doc

The following definition of the SQRT sequence is taken from [43]. Let p now be the

Page 673: genetic_POPULATIONS.doc

ith prime number and z =

Page 674: genetic_POPULATIONS.doc

{x i} is defined by

Page 675: genetic_POPULATIONS.doc

x = (iz) = (iz , ... , iz )mod 1,

Page 676: genetic_POPULATIONS.doc

where themodulo is takencomponent-wise.Despite itssimplicity, the SQRT sequence

Page 677: genetic_POPULATIONS.doc

performs well in [43], where it is applied to a numerical integration problem and com-

Page 678: genetic_POPULATIONS.doc

pared with five quasi random sequences.

Page 679: genetic_POPULATIONS.doc

A.2 Some quasi random sequence generators

Page 680: genetic_POPULATIONS.doc

To describe the structure and the distribution properties of some quasi random se-

Page 681: genetic_POPULATIONS.doc

quences let us define an elementary interval in base b

Page 682: genetic_POPULATIONS.doc

where a , d and b are integers d ≥ 0, 0 ≤ a < b

for i = 1, ... , n. Thus, E is a subinterval of an n-dimensional unit cube I and its ith side has a length of

Page 683: genetic_POPULATIONS.doc

1/b . Figure 11 illustrates 2-dimensional elementary intervals with base b = 2 a nd

Page 684: genetic_POPULATIONS.doc

d = d = 2, and the colored area corresponds to values a = 1 a nd a = 2.

Page 685: genetic_POPULATIONS.doc

Quasi random sequences are called (t, s)-sequences if they satisfy the following

Page 686: genetic_POPULATIONS.doc

distribution property (s is the dimension that we denote by n). For all integers k ≥ 0,

Page 687: genetic_POPULATIONS.doc

the point set {xj} of (t, s)-sequence with kbm ≤ j < ( k + 1)bm has exactly b points

Page 688: genetic_POPULATIONS.doc

on every elementary interval in base b with volume b (in L -metric) [36]. The distribution properties for a (t, s)-sequence are the most preferable, when the dis-

Page 689: genetic_POPULATIONS.doc

tribution parameter t = 0, because then the first bm successive points divide evenly

Page 690: genetic_POPULATIONS.doc

on the corresponding elementary intervals and the same is true for the following b

Page 691: genetic_POPULATIONS.doc

successive points and so on.

Page 692: genetic_POPULATIONS.doc

The class of (t, s)-sequences was introduced by Sobol’ in 1966 (see, [36]). He called

Page 693: genetic_POPULATIONS.doc

them LP -sequences (τ ≡ t) and studied them in base 2. Faure generalized the

Page 694: genetic_POPULATIONS.doc

Fig. 11 Two-dimensional

Page 695: genetic_POPULATIONS.doc

elementary intervals

Page 696: genetic_POPULATIONS.doc

then Van der Corput sequence in base b is the sequence {φ (k)} . Informally, we ∞k=0

may say that in the radical inverse the numbers d , ... , d are reflected using the deci- 1 j

J Glob Optim (2007) 37:405–436

2 2 2

der Corput sequence is a 1-dimensional (t, s)-sequence with t = 0 (see, e.g., [30]).

433

j 1 b

i b

φ (k ) =b d 1

b

d j

b j = (0.d ... d ) , ( 3) 1 j b

b

i b 1 b −1 n

(i)), i = 0, 1, ... ,N − 1 ( 4)

1 n

i b 1 b n (i)), i = 0, 1, ... . ( 5)

i i

p 1

p−1 +· · · +a − 1 z + a . ( 6) p p

Page 697: genetic_POPULATIONS.doc

(t, s)-sequences to arbitrary prime base b ≥ 2 (see, e.g., [36]). Finally, Niederreiter

Page 698: genetic_POPULATIONS.doc

generalized (t, s)-sequence to arbitrary base b ≥ 2.

Page 699: genetic_POPULATIONS.doc

Often a (t, s)-sequence is called the Sobol’ sequence if b = 2 a nd t = 0, and Faure

Page 700: genetic_POPULATIONS.doc

sequence if b is a prime and t = 0 [30]. Sequences that are constructed according to the

Page 701: genetic_POPULATIONS.doc

guidelines given in [30] are called Niederreiter sequences. For a general construction

Page 702: genetic_POPULATIONS.doc

of any (t, s)-sequence, see, e.g., [20, 30].

Page 703: genetic_POPULATIONS.doc

Next, we present some examples how to construct (t, s)-sequences. We omit all

Page 704: genetic_POPULATIONS.doc

implementational details.

Page 705: genetic_POPULATIONS.doc

A.2.1 Van der Corput sequence

Page 706: genetic_POPULATIONS.doc

Let b ≥ 2 be the base and k be an integer. If we write k in base b

k = (d ...d ) ,

where d ∈ { 0, ... , b −1}, i = 1, ... , j, and define a radical inverse function φ by

Page 707: genetic_POPULATIONS.doc

+· · · +

Page 708: genetic_POPULATIONS.doc

mal point as a reflector, for example, if k = 12 = (1100) , t hen φ (12) = 0.0011 .Van

Page 709: genetic_POPULATIONS.doc

A.2.2 Hammersley sequence

Page 710: genetic_POPULATIONS.doc

All but the first component of the n-dimensional Hammersley sequence are van der Corput sequences in an appropriate base. The n-dimensional Hammersley sequence

Page 711: genetic_POPULATIONS.doc

is defined as

Page 712: genetic_POPULATIONS.doc

x = (i/N, φ (i), ... , φ

Page 713: genetic_POPULATIONS.doc

where N is the number of points to be generated and the bases b , ... , b are integers

Page 714: genetic_POPULATIONS.doc

greater than one and pairwise prime.

Page 715: genetic_POPULATIONS.doc

A.2.3 Halton sequence

Page 716: genetic_POPULATIONS.doc

A Hammersley sequence without the first element is called Halton sequence [17]

Page 717: genetic_POPULATIONS.doc

x = (φ (i), ... , φ

Page 718: genetic_POPULATIONS.doc

The advantage over the Hammersley sequence is that a Halton sequence does not

Page 719: genetic_POPULATIONS.doc

require the knowledge of N, the total length of the sequence.

Page 720: genetic_POPULATIONS.doc

A.2.4 Sobol’ sequence

Page 721: genetic_POPULATIONS.doc

We construct the Sobol’ sequence following [4] and [14]. Let us first consider 1-dimensional Sobol’ sequence, for which we need to create a set of direction num- bers v in base 2. To create the direction numbers v , we use coefficients of a primitive

Page 722: genetic_POPULATIONS.doc

polynomial q in the field of binary numbers (compare to Eq. 1 in Sect. 7)

Page 723: genetic_POPULATIONS.doc

q(z) = z +a z

Page 724: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436

3 2 1 b = 1, b = 0, b = 1 a nd b = 1). To generate an n-dimensional sequence, it is

d mod b.

434

i 1

i−1 2

i−2 p

i−p v i−p

⊕2 p

, i > p.

i 1

1 2

2 3

3

4 3 2 1

b c 0

b c −1 j

b j ,

b 1 j b

c m

=

j

n=m

h!h

i b b

n−1 (φ (i))), b

(1994)

Page 725: genetic_POPULATIONS.doc

Then, employing the bitwise binary exclusive-or operator ⊕ we derive the direction

Page 726: genetic_POPULATIONS.doc

numbers from the formula

Page 727: genetic_POPULATIONS.doc

v = a v ⊕a v ··⊕ · ⊕a v

Page 728: genetic_POPULATIONS.doc

The ith recurrence in 1-dimensional Sobol’ sequence is now formed as

Page 729: genetic_POPULATIONS.doc

x = b v ⊕b v ⊕b v ··⊕ · ,

Page 730: genetic_POPULATIONS.doc

where ··· b b b is the binary representation of i (for example, if i = 1011, then

sufficient to choose n different primitive polynomials and calculate n different sets of

Page 731: genetic_POPULATIONS.doc

direction numbers. For more details, see [4].

Page 732: genetic_POPULATIONS.doc

A.2.5 Faure sequence

Page 733: genetic_POPULATIONS.doc

The followingexample of Faure sequence ismainly from [43].Let us define a function g

Page 734: genetic_POPULATIONS.doc

g(φ (k)) =+· · · +

Page 735: genetic_POPULATIONS.doc

where b is the base, φ (k) = (0.d ... d ) is an element of the Van der Corput

Page 736: genetic_POPULATIONS.doc

sequence and

Page 737: genetic_POPULATIONS.doc

m(h −m) !

Page 738: genetic_POPULATIONS.doc

Now, an n-dimensional (t, s)-sequence is defined as

Page 739: genetic_POPULATIONS.doc

x = (φ (i), g(φ (i)), ... , g

Page 740: genetic_POPULATIONS.doc

and a Faure sequence is obtained when b is the smallest prime larger or equal to the

Page 741: genetic_POPULATIONS.doc

dimension n [36].

Page 742: genetic_POPULATIONS.doc

References

Page 743: genetic_POPULATIONS.doc

1. Acworth, P., Broadie, M., Glasserman, P.: A comparison of some Monte Carlo and quasi Monte

Page 744: genetic_POPULATIONS.doc

Carlo techniques for option pricing. In: Niederreiter, H., Hellekalek, P., Larcher, G., Zinterhof,

Page 745: genetic_POPULATIONS.doc

P. (eds). Monte Carlo and Quasi-Monte Carlo Methods 1996, number 127 in Lecture notes in

Page 746: genetic_POPULATIONS.doc

statistics, pp. 1–18. Springer-Verlag, New York (1998)

Page 747: genetic_POPULATIONS.doc

2. Ali, M.M., Storey, C.: Modified controlled random search algorithms. Int. J. Comput. Mathemat.

Page 748: genetic_POPULATIONS.doc

53, 229–235 (1994)

Page 749: genetic_POPULATIONS.doc

3. Ali,M.M., Storey, C.: Topographical multilevel single linkage. J. Global Optimizat. 5(4), 349–358

Page 750: genetic_POPULATIONS.doc

4. Bratley, P., Fox, B.L.: Algorithm 659: implementing Sobol’s quasirandom sequence generator.

Page 751: genetic_POPULATIONS.doc

ACM Trans. Mathemat. Software 14(1), 88–100 (1988)

Page 752: genetic_POPULATIONS.doc

5. Bratley, P., Fox, B.L., Niederreiter, H.: Implementation and tests of low-discrepancy sequences.

Page 753: genetic_POPULATIONS.doc

ACM Trans. Model. Comput. Simulat. 2(3), 195–213 (1992)

Page 754: genetic_POPULATIONS.doc

6. Couture, R., L’Ecuyer, P.: Distribution properties of multiply-with-carry random number

Page 755: genetic_POPULATIONS.doc

generators. Mathemat. Computat. 66(218), 591–607 (1997)

Page 756: genetic_POPULATIONS.doc

7. Dekkers, A., Aarts, E.: Global optimization and simulated annealing.Mathemat. Program. 50(3),

Page 757: genetic_POPULATIONS.doc

367–393 (1991)

Page 758: genetic_POPULATIONS.doc

8. DIEHARD: a battery of tests for random number generators developed by George Marsaglia.

Page 759: genetic_POPULATIONS.doc

http://stat.fsu.edu/~geo/diehard.html , September (2004).

Page 760: genetic_POPULATIONS.doc

9. Diggle, P.J.: Statistical Analysis of Spatial Point Patterns. Academic Press, London (1983)

Page 761: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 435

(2004)

(1998)

(2002)

(1989)

(1998)

Page 762: genetic_POPULATIONS.doc

10. Dykes, S., Rosen, B.: Parallel very fast simulated reannealing by temperature block partitioning.

Page 763: genetic_POPULATIONS.doc

In Proceedings of the 1994 IEEE International Conference on Systems, Man, and Cybernetics,

Page 764: genetic_POPULATIONS.doc

vol. 2, pp. 1914–1919 (1994)

Page 765: genetic_POPULATIONS.doc

11. Eiben,A.E.,Hinterding,R.,Michalewicz,Z.:Parameter control in evolutionaryalgorithms. IEEE

Page 766: genetic_POPULATIONS.doc

Trans. Evol. Computat. 3(2), 124–141 (1999)

Page 767: genetic_POPULATIONS.doc

12. Floudas, C.A., Pardalos, P.M. (eds.): RecentAdvances inGlobalOptimization. PrincetonUniver-

Page 768: genetic_POPULATIONS.doc

sity Press, Princeton, NJ (1992)

Page 769: genetic_POPULATIONS.doc

13. Genetic and evolutionary algorithm toolbox for use with Matlab. http://www.geatbx.com/, June

Page 770: genetic_POPULATIONS.doc

14. Gentle, J.E.:RandomNumberGeneration andMonteCarloMethods. Springer-Verlag,NewYork

Page 771: genetic_POPULATIONS.doc

15. Glover, F.,Kochenberger,G.A. (eds.):Handbook ofMetaheuristics.KluwerAcademic Publishers,

Page 772: genetic_POPULATIONS.doc

Boston (2003)

Page 773: genetic_POPULATIONS.doc

16. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-

Page 774: genetic_POPULATIONS.doc

Wesley, New York (1989)

Page 775: genetic_POPULATIONS.doc

17. Halton, J.H.: On the efficiency of certain quasi-random sequences of points in evaluating multi-

Page 776: genetic_POPULATIONS.doc

dimentional integrals. Numerische Mathematik, 2, 84–90 (1960)

Page 777: genetic_POPULATIONS.doc

18. Horst,R., Pardalos, P.M. (eds.):Handbook ofGlobalOptimization.KluwerAcademic Publishers,

Page 778: genetic_POPULATIONS.doc

Dordrecht (1995)

Page 779: genetic_POPULATIONS.doc

19. Kingsley, M.C.S., Kanneworff, P., Carlsson, D.M.: Buffered random sampling: a sequential inhib-

Page 780: genetic_POPULATIONS.doc

ited spatial point process applied to sampling in a trawl survey for northern shrimp pandalus

Page 781: genetic_POPULATIONS.doc

borealis in west Greenland waters. ICES J. Mar. Sci. 61(1), 12–24 (2004)

Page 782: genetic_POPULATIONS.doc

20. Larcher, G.: Digital point sets: analysis and applications. In: Hellekalek, P., Larcher, G. (eds.)

Page 783: genetic_POPULATIONS.doc

Random and Quasi Random Point Sets, number 138 in Lecture Notes in Statistics, pp. 167–222.

Page 784: genetic_POPULATIONS.doc

Springer-Verlag, New York (1998)

Page 785: genetic_POPULATIONS.doc

21. Lei,G.:Adaptive random search inquasi-Monte Carlomethods for globaloptimization.Comput.

Page 786: genetic_POPULATIONS.doc

Mathemat. Appl. 43(6–7), 747–754 (2002)

Page 787: genetic_POPULATIONS.doc

22. Lemieux,C., L’Ecuyer, P.:On selection criteria for lattice rules and otherquasi-Monte Carlo point

Page 788: genetic_POPULATIONS.doc

sets. Mathemat. Comput. Simulat. 55(1–3), 139–148 (2001)

Page 789: genetic_POPULATIONS.doc

23. Maaranen, H.: On Global Optimization with Aspects toMethod Comparison and Hybridization.

Page 790: genetic_POPULATIONS.doc

Licentiate thesis, Department of Mathematical Information Technology, University of Jyväskylä

Page 791: genetic_POPULATIONS.doc

24. Maaranen, H., Miettinen, K., Mäkelä, M.M.: Quasi-random initial population for genetic algo-

Page 792: genetic_POPULATIONS.doc

rithms. Comput. Mathemat. Appl. 47(12), 1885–1895 (2004)

Page 793: genetic_POPULATIONS.doc

25. Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs. Springer-Verlag,

Page 794: genetic_POPULATIONS.doc

Berlin (1994)

Page 795: genetic_POPULATIONS.doc

26. Michalewicz, Z., Logan, T.D., Swaminathan, S.: Evolutionary operators for continuous convex

Page 796: genetic_POPULATIONS.doc

parameter spaces. In: Sebald, A.V., Fogel, L.J. (eds.) Proceedings of the 3rd Annual Conference

Page 797: genetic_POPULATIONS.doc

on Evolutionary Programming, pp. 84–97.World Scientific Publishing, River Edge, NJ (1994)

Page 798: genetic_POPULATIONS.doc

27. Morokoff, W.J., Caflisch, R.E.: Quasi-random sequences and their discrepancies. SIAM J. Sci.

Page 799: genetic_POPULATIONS.doc

Comput. 15(6), 1251–1279 (1994)

Page 800: genetic_POPULATIONS.doc

28. Neter, J.,Wasserman,W.,Whitmore,G.A.:Applied Statistics.Allyn andBacon, Inc.,Boston, third

Page 801: genetic_POPULATIONS.doc

edition (1988)

Page 802: genetic_POPULATIONS.doc

29. Niederreiter, H.: Quasi-Monte Carlo methods for global optimization. In: Grossmann,W., Pflug,

Page 803: genetic_POPULATIONS.doc

G., Vincze, I., Wertz, W. (eds.) Proceedings of the 4th Pannonian Symposium on Mathematical

Page 804: genetic_POPULATIONS.doc

Statistics, pp. 251–267 (1983)

Page 805: genetic_POPULATIONS.doc

30. Niederreiter, H.: Random Number Generation and Quasi-Monte Carlo Methods. SIAM, Phila-

Page 806: genetic_POPULATIONS.doc

delphia (1992)

Page 807: genetic_POPULATIONS.doc

31. Nix, A.E., Vose, M.D.: Modeling genetic algorithms with Markov chains. Ann. Mathemat. Arti.

Page 808: genetic_POPULATIONS.doc

Intell. 5(1), 79–88 (1992)

Page 809: genetic_POPULATIONS.doc

32. Pardalos, P.M.,Romeijn,H.E. (eds.):Handbook ofGlobalOptimization, vol. 2.KluwerAcademic

Page 810: genetic_POPULATIONS.doc

Publishers, Boston (2002)

Page 811: genetic_POPULATIONS.doc

33. Press, W.H., Teukolsky, S.A.: Quasi- (that is, sub-) random numbers. Comput. Phys. 3(6), 76–79

Page 812: genetic_POPULATIONS.doc

34. Ripley, B.D.: Spatial Statistics. John Wiley & Sons, New York (1981)

Page 813: genetic_POPULATIONS.doc

35. Snyder, W.C.: Accuracy estimation for quasi-Monte Carlo simulations. Mathemat Comput. Sim-

Page 814: genetic_POPULATIONS.doc

ulat. 54(1–3), 131–143 (2000)

Page 815: genetic_POPULATIONS.doc

36. Sobol’, I.M.: On quasi-Monte Carlo integrations. Mathemat. Comput. Simulat. 47(2–5), 103–112

37. Sobol’, I.M., Bakin, S.G.: On the crude multidimensional search. J. Computat. Appl. Mathemat.

Page 816: genetic_POPULATIONS.doc

56(3), 283–293 (1994)

Page 817: genetic_POPULATIONS.doc

J Glob Optim (2007) 37:405–436 436

2

Page 818: genetic_POPULATIONS.doc

38. Sobol, I.M., Tutunnikov,A.V.:A varience reducingmultiplier forMonteCarlo integrations.Math-

Page 819: genetic_POPULATIONS.doc

emat. Comput. Model. 23(8–9), 87–96 (1996)

Page 820: genetic_POPULATIONS.doc

39. Tang,H.-C.:Using an adaptive geneticalgorithmwith reversals to find good second-ordermultiple

Page 821: genetic_POPULATIONS.doc

recursive random number generators. Mathemat. Methods Operat. Res. 57(1), 41–48 (2003)

Page 822: genetic_POPULATIONS.doc

40. Testproblems for globaloptimization. http://www.imm.dtu.dk/~km/GlobOpt/testex/ , June (2004)

Page 823: genetic_POPULATIONS.doc

41. Test problems in R . http://iridia.ulb.ac.be/~aroli/ICEO/Functions/Functions.html , June (2004)

Page 824: genetic_POPULATIONS.doc

42. Trafalis, B., Kasap, S.: A novel metaheuristics approach for continuous global optimization.

Page 825: genetic_POPULATIONS.doc

J. Global Optimizat. 23(2), 171–190 (2002)

Page 826: genetic_POPULATIONS.doc

43. Tuffin, B.: On the use of low discrepancy sequences inMonte Carlo methods.Monte CarloMeth-

Page 827: genetic_POPULATIONS.doc

ods Appl. 2(4), 295–320 (1996)

Page 828: genetic_POPULATIONS.doc

44. Weisstein, E.W.: “circle packing”. From MathWorld–A Wolfram Web Resource. http://math-

Page 829: genetic_POPULATIONS.doc

world.wolfram.com/CirclePacking.html, September (2004)

Page 830: genetic_POPULATIONS.doc

45. Wikramaratna, R.S.: ACORN—a new method for generating sequences of uniformly distributed

Page 831: genetic_POPULATIONS.doc

pseudo-random numbers. J. Computat. Phys. 83(1), 16–31 (1989)

Page 832: genetic_POPULATIONS.doc

46. Wright, A.H., Zhao, Y.: Markov chain models of genetic algorithms. In: Banzhaf, W., Daida, J.,

Page 833: genetic_POPULATIONS.doc

Eiben, A.E., Garzon, M.H., Honavar, V., Jakiela, M., Smith, R.E. (eds.) GECCO-99: Proceed- ings of the Genetic and Evolutionary Computation Conference, pp. 734–741.Morgan Kauffman,

Page 834: genetic_POPULATIONS.doc

SanMateo, California (1999)